{"schema_version":"onlylabs.public_analyses.v1","title":"onlylabs Frontier lab analysis export","description":"Structured public onlylabs agent analyses: generated markdown reports, cited evidence, provenance, and stable report URLs.","url":"https://onlylabs.fyi/analysis?category=frontier-lab","json_url":"https://onlylabs.fyi/analysis.json?category=frontier-lab","generated_at":"2026-06-11T16:54:24.380Z","scope":{"category":"frontier-lab","label":"Frontier lab"},"count":18,"analyses":[{"org_slug":"amazon","url":"https://onlylabs.fyi/analysis/amazon","json_url":"https://onlylabs.fyi/analysis/amazon/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/amazon/evidence.json","dossier_url":"https://onlylabs.fyi/labs/amazon","org":{"slug":"amazon","name":"Amazon (Nova)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://nova.amazon.com/"},"title":"Amazon (Nova) analysis","summary":"Amazon is betting its AI identity on agentic AI as the organizing principle. The evidence pack shows a coordinated push: multiple amazon.science posts on agent design patterns, open-source agent frameworks spanning RL training, multi-agent evolution, and compliance verification, and product launches including Nova Act and the perception agent harness. Alongside this, Amazon is building the trust infrastructure…","markdown":"```json\n{\n  \"content\": \"## Thesis\\n\\nAmazon is betting its AI identity on agentic AI as the organizing principle. The evidence pack shows a coordinated push: multiple amazon.science posts on agent design patterns [P1](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems)[P2](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai)[P10](https://www.amazon.science/blog/designing-ai-agents-that-know-when-to-step-back)[P11](https://www.amazon.science/blog/how-agentic-ai-helps-heal-the-systems-we-cant-replace)[P16](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale), open-source agent frameworks spanning RL training, multi-agent evolution, and compliance verification [E5](https://github.com/amazon-science/reskill)[E12](https://github.com/amazon-science/EvoMAS)[E24](https://github.com/amazon-science/compagent)[E55](https://github.com/amazon-science/agentic-forking-path), and product launches including Nova Act [W4](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents) and the perception agent harness [W1](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source). Alongside this, Amazon is building the trust infrastructure needed for agent deployment at scale — formal verification [P13](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum)[P19](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine)[E32](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine), post-quantum cryptography [P15](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon)[E42](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon), privacy-preserving training with cryptographic defenses [P21](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data)[E30](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data), and LLM catastrophic risk certification [P20](https://www.amazon.science/blog/how-catastrophic-is-your-llm)[E31](https://www.amazon.science/blog/how-catastrophic-is-your-llm). The Nova model family serves both as a standalone offering (Chronos-2 at 12.5M downloads [E1](https://huggingface.co/amazon/chronos-2)) and a customization platform (Nova Forge hyperparameter optimization [W2](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/), Nova for molecular-property prediction [P18](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery)[E35](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery)). Amazon leverages its operational DNA — supply chain optimization [P23](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration)[P24](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network)[E22](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network), datacenter network innovation [P5](https://arxiv.org/pdf/2604.15261)[E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks), and security at scale [P16](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale)[E41](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale) — as its differentiation. Critically, **this pack contains zero hiring signals**, making workforce strategy an information gap.\\n\\n## Signal desks\\n\\n### Hiring\\n\\nNo cited evidence in this pack. No job listings, career pages, or role announcements appear across the 28 pages, 60 events, or 4 web search results.\\n\\n### Forks\\n\\nNo cited evidence in this pack. All GitHub activity consists of first-party repos published by `amazon-science`; no forked upstream repositories are identified in the evidence [E5](https://github.com/amazon-science/reskill)[E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl)[E12](https://github.com/amazon-science/EvoMAS)[E17](https://github.com/amazon-science/adaptive-layerwise-perturbation)[E19](https://github.com/amazon-science/temporal-reasoning-dataset)[E20](https://github.com/amazon-science/PROF-GRPO)[E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus)[E23](https://github.com/amazon-science/RecArena)[E24](https://github.com/amazon-science/compagent)[E25](https://github.com/amazon-science/SWAN)[E28](https://github.com/amazon-science/rmir)[E33](https://github.com/amazon-science/expert-upcycling)[E34](https://github.com/amazon-science/CodeStruct)[E39](https://github.com/amazon-science/TransitionFlowMatching)[E51](https://github.com/amazon-science/storm-referring-multi-object-grounding)[E54](https://github.com/amazon-science/acclaim)[E55](https://github.com/amazon-science/agentic-forking-path)[E57](https://github.com/amazon-science/papercode-coordinating-spot-and-contracts)[E60](https://github.com/amazon-science/TSFM-Biases).\\n\\n### Releases\\n\\n- **Chronos-2**: Time-series forecasting foundation model, 119M params, Apache 2.0, 12.5M HuggingFace downloads, 317 likes — the highest-traction artifact in the pack [E1](https://huggingface.co/amazon/chronos-2).\\n- **P-EAGLE speculative decoding series**: Long-context models (gpt-oss-20b/120b, Qwen3-Coder-30B) targeting inference efficiency via speculative decoding, all Apache 2.0 [E13](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E14](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E15](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context)[E56](https://huggingface.co/amazon/gpt-oss-120b-p-eagle).\\n- **HQwen3 primed fine-tune batch**: At least 10 models released 2026-03-31 using GKA, GDN, Mamba2, and BMOJOF priming methods on Qwen3 8B/32B backbones across Instruct and Reasoner variants, Apache 2.0 [E38](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)[E40](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct)[E43](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner)[E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner)[E45](https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct)[E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct)[E48](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct)[E49](https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct)[E50](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner).\\n- **Agent infrastructure**: `reskill` — veRL extension for agent RL training with skill co-evolution [E5](https://github.com/amazon-science/reskill); `EvoMAS` — evolutionary multi-agent system generation, ICML 2026 [E12](https://github.com/amazon-science/EvoMAS); `CompAgent` — visual compliance verification agent [E24](https://github.com/amazon-science/compagent); `agentic-forking-path` [E55](https://github.com/amazon-science/agentic-forking-path).\\n- **Evaluation and dataset repos**: `temporal-reasoning-dataset` — multilingual temporal reasoning benchmark [E19](https://github.com/amazon-science/temporal-reasoning-dataset); `hallucination-benchmark-trivialplus` — ACL 2026 long-context hallucination detection benchmark [E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus); `RMIR` — reasoning-intensive multimodal image retrieval benchmark [E28](https://github.com/amazon-science/rmir); `RecArena` [E23](https://github.com/amazon-science/RecArena).\\n- **Training infrastructure**: `dualkv-flash-attn-for-rl` — shared-prompt flash attention for efficient RL training [E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl); `PROF-GRPO` [E20](https://github.com/amazon-science/PROF-GRPO); `expert-upcycling` (14 GitHub stars) [E33](https://github.com/amazon-science/expert-upcycling); `adaptive-layerwise-perturbation` [E17](https://github.com/amazon-science/adaptive-layerwise-perturbation).\\n- **Library releases with active cadence**: `concurry` v0.13.1/v0.13.2 [E10](https://github.com/amazon-science/concurry/releases/tag/v0.13.2)[E11](https://github.com/amazon-science/concurry/releases/tag/v0.13.1); `azcausal` v0.2.4.3/v0.2.5 [E29](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E52](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3); `uniqsketch` v1.2.1 [E46](https://github.com/amazon-science/uniqsketch/releases/tag/v1.2.1).\\n- **Other notable repos**: `TransitionFlowMatching` — AISTATS 2026, image/video generation, 12 stars [E39](https://github.com/amazon-science/TransitionFlowMatching); `SWAN` — semantic watermarking, ACL 2026 [E25](https://github.com/amazon-science/SWAN); `CodeStruct` [E34](https://github.com/amazon-science/CodeStruct); `TSFM-Biases` — time-series foundation model bias analysis [E60](https://github.com/amazon-science/TSFM-Biases); `storm-referring-multi-object-grounding` [E51](https://github.com/amazon-science/storm-referring-multi-object-grounding); `acclaim` [E54](https://github.com/amazon-science/acclaim); `papercode-coordinating-spot-and-contracts` [E57](https://github.com/amazon-science/papercode-coordinating-spot-and-contracts).\\n\\n### Talking\\n\\n- **Agentic AI is the dominant narrative**: Posts cover bridging intent and execution in agentic systems [P1](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems)[E4](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems), four approaches to real-world grounding for AI agents [P2](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai)[E3](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai), UX design for human-AI coordination in agentic systems [P10](https://www.amazon.science/blog/designing-ai-agents-that-know-when-to-step-back), agentic AI for healing legacy systems that can't be replaced [P11](https://www.amazon.science/blog/how-agentic-ai-helps-heal-the-systems-we-cant-replace), RuleForge agentic vulnerability detection producing rules 336% faster [P16](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale)[E41](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale), the open-source perception agent harness [W1](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source), and Amazon's overall agentic-AI approach with Nova Act training model capabilities, orchestration, and tool controls as one integrated system [W4](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents).\\n- **Trust, safety, and formal verification stack**: Amazon's responsible-AI pipeline embedding safety throughout the development lifecycle [P22](https://www.amazon.science/blog/building-trust-into-ai)[E27](https://www.amazon.science/blog/building-trust-into-ai); statistical framework for certifying LLM catastrophic failure likelihood in adversarial conversations [P20](https://www.amazon.science/blog/how-catastrophic-is-your-llm)[E31](https://www.amazon.science/blog/how-catastrophic-is-your-llm); reproducing training-data extraction attacks and cryptographic defenses that stop them [P21](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data)[E30](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data); formally verified AES-XTS as first AES algorithm in s2n-bignum [P13](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum)[E58](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum); verifying and optimizing post-quantum cryptography with automated reasoning [P15](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon)[E42](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon); Isabelle/HOL proof assistant enabling the world's first formally verified cloud hypervisor (Nitro Isolation Engine) [P19](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine)[E32](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine); academic collaboration delivering real-world security to customers [P6](https://www.amazon.science/news/how-academic-collaboration-delivers-real-world-security-to-amazon-customers).\\n- **Inference and training efficiency**: New scaling law connecting architectural choices to loss, identifying models with up to 47% throughput improvement at no accuracy loss [P26](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy)[E16](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy); thesis that intelligence isn't about parameter count but inference time — larger models become less insightful without reduced inference time [P8](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time)[E37](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time); LoRA target module selection ablation study on accuracy-efficiency trade-offs [P12](https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning)[E59](https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning); Promptimus automated prompt-engineering framework for improving prompts without manual work [P25](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering)[E18](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering); training LLMs to generate diverse accurate reasoning paths using global forking tokens [P27](https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions)[E9](https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions).\\n- **Operational optimization**: Mechanism design theory applied to Amazon-vendor supply chain collaboration without disclosing private information [P23](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration)[E26](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration); new tools for optimizing middle-mile delivery networks under uncertainty [P24](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network)[E22](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network); RNG flat datacenter networks using quasi-random graphs and ShuffleBox optical devices, now default for most AWS workloads, up to 45% cheaper than fat trees [P5](https://arxiv.org/pdf/2604.15261)[E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks); 12-year-old forecasting paper still proving durable [P7](https://www.amazon.science/blog/why-a-12-year-old-forecasting-paper-has-stood-the-test-of-time).\\n- **Domain applications**: Customized Amazon Nova models unifying molecular-property prediction in drug discovery, serving as reasoning partner for medical chemists [P18](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery)[E35](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery); AWS–Johns Hopkins antibody developability benchmark with diverse public antibody datasets for AI-guided antibody design [P17](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design)[E36](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design); LLM-based TTS improvements via LoRA, data augmentation, and chain-of-thought reasoning for accent-free polyglot output [P14](https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems)[E53](https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems); AI changing the nature of mathematical research [P9](https://www.amazon.science/blog/how-ai-is-changing-the-nature-of-mathematical-research).\\n- **Data and evaluation**: Ground truth framed as a process, not a dataset — challenges in auto-fact-checking long AI-generated research reports [E6](https://www.amazon.science/blog/ground-truth-is-a-process-not-a-dataset); Nova Sonic Test Harness for evaluating voice agents at scale with audio-hallucination detection and LLM-as-judge [W3](https://aws.amazon.com/blogs/machine-learning/evaluate-your-amazon-nova-sonic-voice-agent-at-scale-no-microphone-required/); hyperparameter optimization on Amazon Nova Forge covering data mixing, learning rate, checkpoint selection [W2](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/); Amazon Research Awards funding recipients across 49 universities in 11 countries with access to Amazon public datasets and AWS AI/ML services [P28](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced)[E8](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced).\\n\\n## Shipping\\n\\nAmazon ships across four lanes in this evidence window:\\n\\n1. **Models**: Chronos-2 dominates with 12.5M downloads [E1](https://huggingface.co/amazon/chronos-2); P-EAGLE speculative decoding series across GPT-OSS and Qwen3-Coder backbones [E13](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E14](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E15](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context)[E56](https://huggingface.co/amazon/gpt-oss-120b-p-eagle); a large batch of primed HQwen3 fine-tunes using GKA, GDN, Mamba2, and BMOJOF methods [E38](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)[E40](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct)[E43](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner)[E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner)[E45](https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct)[E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct)[E48](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct)[E49](https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct)[E50](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner).\\n2. **Agent frameworks**: `reskill` for agent RL with skill co-evolution [E5](https://github.com/amazon-science/reskill); `EvoMAS` for evolutionary multi-agent systems [E12](https://github.com/amazon-science/EvoMAS); `CompAgent` for visual compliance [E24](https://github.com/amazon-science/compagent); perception agent harness with annotation and verification primitives [W1](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source); Nova Act as an integrated agent-building service [W4](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents).\\n3. **Evaluation infrastructure**: Hallucination detection benchmarks [E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus), temporal reasoning datasets [E19](https://github.com/amazon-science/temporal-reasoning-dataset), multimodal retrieval benchmarks [E28](https://github.com/amazon-science/rmir), Nova Sonic test harness with audio-hallucination detection [W3](https://aws.amazon.com/blogs/machine-learning/evaluate-your-amazon-nova-sonic-voice-agent-at-scale-no-microphone-required/), Antibody Developability Benchmark [E36](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design).\\n4. **Core infrastructure**: RNG flat datacenter networks now default for most AWS workloads, up to 45% cheaper [P5](https://arxiv.org/pdf/2604.15261)[E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks); formally verified AES-XTS in s2n-bignum [P13](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum)[E58](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum); `concurry`, `azcausal`, and `uniqsketch` library releases on active cadences [E10](https://github.com/amazon-science/concurry/releases/tag/v0.13.2)[E11](https://github.com/amazon-science/concurry/releases/tag/v0.13.1)[E29](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E46](https://github.com/amazon-science/uniqsketch/releases/tag/v1.2.1)[E52](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3).\\n\\n## Research themes\\n\\n- **Agentic AI systems**: Design patterns for intent-execution bridging [P1](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems)[E4](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems), real-world grounding approaches [P2](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai)[E3](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai), human-AI coordination UX [P10](https://www.amazon.science/blog/designing-ai-agents-that-know-when-to-step-back), legacy-system integration [P11](https://www.amazon.science/blog/how-agentic-ai-helps-heal-the-systems-we-cant-replace), multi-agent evolutionary systems [E12](https://github.com/amazon-science/EvoMAS), and agentic forking-path architectures [E55](https://github.com/amazon-science/agentic-forking-path).\\n- **Trustworthy AI stack**: Formal verification with Isabelle/HOL for cloud hypervisors [P19](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine)[E32](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine); verified AES-XTS and post-quantum cryptography [P13](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum)[P15](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon)[E42](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon)[E58](https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum); responsible-AI pipeline development [P22](https://www.amazon.science/blog/building-trust-into-ai)[E27](https://www.amazon.science/blog/building-trust-into-ai); LLM catastrophic risk certification through statistical frameworks [P20](https://www.amazon.science/blog/how-catastrophic-is-your-llm)[E31](https://www.amazon.science/blog/how-catastrophic-is-your-llm); cryptographic defenses against training-data extraction [P21](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data)[E30](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data); semantic watermarking [E25](https://github.com/amazon-science/SWAN).\\n- **Efficient training and inference**: Scaling laws linking architecture to inference throughput [P26](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy)[E16](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy); speculative decoding via P-EAGLE [E13](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E14](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E15](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context); LoRA target module optimization [P12](https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning)[E59](https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning); RL training efficiency with DualKV flash attention [E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl); expert upcycling [E33](https://github.com/amazon-science/expert-upcycling); inference-time intelligence thesis [P8](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time)[E37](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time); diverse reasoning trace training [P27](https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions)[E9](https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions).\\n- **Domain-specific AI**: Time-series forecasting via Chronos-2 [E1](https://huggingface.co/amazon/chronos-2); drug discovery with customized Nova models [P18](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery)[E35](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery); antibody design benchmarking [P17](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design)[E36](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design); LLM-based text-to-speech quality and robustness [P14](https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems)[E53](https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems); mechanism design for supply chain [P23](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration)[E26](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration).\\n- **Systems and optimization science**: Quasi-random graph datacenter networks [P5](https://arxiv.org/pdf/2604.15261)[E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks); middle-mile logistics under uncertainty [P24](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network)[E22](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network); forecasting methodology [P7](https://www.amazon.science/blog/why-a-12-year-old-forecasting-paper-has-stood-the-test-of-time); automated prompt engineering [P25](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering)[E18](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering); causal inference libraries [E29](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E52](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3).\\n\\n## Hiring & scaling\\n\\nNo hiring signals appear in this evidence pack. The pattern of 30+ open-source repos from `amazon-science` and sustained blog output from `amazon.science` suggests an active, publishing research organization [E5](https://github.com/amazon-science/reskill)[E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl)[E12](https://github.com/amazon-science/EvoMAS)[E17](https://github.com/amazon-science/adaptive-layerwise-perturbation)[E19](https://github.com/amazon-science/temporal-reasoning-dataset)[E20](https://github.com/amazon-science/PROF-GRPO)[E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus)[E23](https://github.com/amazon-science/RecArena)[E24](https://github.com/amazon-science/compagent)[E25](https://github.com/amazon-science/SWAN)[E28](https://github.com/amazon-science/rmir)[E33](https://github.com/amazon-science/expert-upcycling)[E34](https://github.com/amazon-science/CodeStruct)[E39](https://github.com/amazon-science/TransitionFlowMatching)[E51](https://github.com/amazon-science/storm-referring-multi-object-grounding)[E54](https://github.com/amazon-science/acclaim)[E55](https://github.com/amazon-science/agentic-forking-path)[E57](https://github.com/amazon-science/papercode-coordinating-spot-and-contracts)[E60](https://github.com/amazon-science/TSFM-Biases), but roles, locations, team sizes, headcount growth, and geographic hubs cannot be estimated from the supplied evidence. The Amazon Research Awards program engages 49 universities across 11 countries [P28](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced)[E8](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced), which may serve as an academic pipeline, but no conversion data into direct hiring is available. This is a notable gap for operators tracking Amazon's AI workforce buildout.\\n\\n## Data-business implications\\n\\n- **Data demand**: Chronos-2's 12.5M downloads signal strong enterprise appetite for time-series foundation models — an opportunity for curated forecasting dataset products [E1](https://huggingface.co/amazon/chronos-2). The Antibody Developability Benchmark [E36](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design) and hallucination-benchmark-trivialplus [E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus) create new structured evaluation datasets; the temporal-reasoning-dataset spans multilingual benchmarks [E19](https://github.com/amazon-science/temporal-reasoning-dataset); RMIR extends to multimodal retrieval evaluation [E28](https://github.com/amazon-science/rmir). Amazon Research Awards grant 49 universities access to Amazon public datasets, expanding the data ecosystem [E8](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced). Nova Forge's data mixing capability blends customer training data with curated datasets to prevent catastrophic forgetting during domain customization [W2](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/).\\n- **Evals and quality**: Nova Sonic Test Harness introduces audio-hallucination detection and LLM-as-judge evaluation at scale for voice agents [W3](https://aws.amazon.com/blogs/machine-learning/evaluate-your-amazon-nova-sonic-voice-agent-at-scale-no-microphone-required/). The \\\"ground truth is a process\\\" framing signals evolving eval methodologies beyond static benchmarks [E6](https://www.amazon.science/blog/ground-truth-is-a-process-not-a-dataset). The hallucination detection benchmark explicitly targets long-context RAG-based evaluation [E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus). These create tooling opportunities for automated quality pipelines.\\n- **Infrastructure**: RNG flat networks — now default for most AWS workloads and up to 45% cheaper than fat trees — represent a datacenter topology shift with implications for training and inference cluster design [P5](https://arxiv.org/pdf/2604.15261)[E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks). The scaling law connecting architecture to inference throughput (47% improvement) [P26](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy)[E16](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy) and DualKV flash attention for RL training with large rollouts [E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl) point to specialized infrastructure needs for RL-based agent training at scale. `reskill` extends veRL for agent RL with skill co-evolution [E5](https://github.com/amazon-science/reskill).\\n- **Tooling**: Nova Forge's hyperparameter optimization framework addresses learning rate, data mixing ratio, checkpoint selection, and training technique interactions [W2](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/). Promptimus provides automated prompt engineering by targeting specific failure points [E18](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering). `concurry` [E10](https://github.com/amazon-science/concurry/releases/tag/v0.13.2)[E11](https://github.com/amazon-science/concurry/releases/tag/v0.13.1) and `azcausal` [E29](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E52](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3) are utility libraries with active release cadences suitable for integration into data and ML platform toolchains.\\n- **Safety and deployment**: The responsible-AI pipeline [P22](https://www.amazon.science/blog/building-trust-into-ai)[E27](https://www.amazon.science/blog/building-trust-into-ai), LLM catastrophic risk certification framework [P20](https://www.amazon.science/blog/how-catastrophic-is-your-llm)[E31](https://www.amazon.science/blog/how-catastrophic-is-your-llm), privacy-preserving training with cryptographic defenses [P21](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data)[E30](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data), and CompAgent for visual compliance verification [E24](https://github.com/amazon-science/compagent) create safety tooling and governance infrastructure opportunities. Nova Act's reliability-first design, training model capabilities and orchestration together as one integrated system [W4](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents), and the perception agent harness [W1](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source) position agentic AI as a product surface requiring new monitoring, evaluation, and guardrail infrastructure.\\n- **Deployment optimization**: P-EAGLE speculative decoding models across parameter scales (20B to 120B) [E13](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E14](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E15](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context)[E56](https://huggingface.co/amazon/gpt-oss-120b-p-eagle) and the inference-time intelligence thesis [E37](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time) indicate deployment optimization around latency-sensitive agent workloads. The HQwen3 primed-series using GKA, GDN, Mamba2, and BMOJOF priming methods [E38](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)[E40](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct)[E43](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner)[E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner)[E45](https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct)[E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct)[E48](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct)[E49](https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct)[E50](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner) reflects systematic exploration of efficient deployment architectures that could inform serving infrastructure decisions.\\n\\n## Traction highlights\\n\\n- **Chronos-2**: 12.5M HuggingFace downloads, 317 likes — the standout traction artifact [E1](https://huggingface.co/amazon/chronos-2).\\n- **GKA-primed-HQwen3-32B-Instruct**: 61,931 downloads [E38](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct).\\n- **GKA-primed-HQwen3-8B-Reasoner**: 3,941 downloads [E43](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner).\\n- **GKA-primed-HQwen3-8B-Instruct**: 3,241 downloads [E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct).\\n- **GDN-primed-HQwen3-8B-Instruct**: 1,339 downloads [E48](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct).\\n- **expert-upcycling**: 14 GitHub stars [E33](https://github.com/amazon-science/expert-upcycling).\\n- **TransitionFlowMatching**: 12 GitHub stars [E39](https://github.com/amazon-science/TransitionFlowMatching).\\n- **reskill**: 5 GitHub stars [E5](https://github.com/amazon-science/reskill).\\n- **HN engagement modest**: RNG flat networks post drew 4 points/2 comments [E2](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks); inference-time intelligence post drew 3 points/0 comments [E37](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time).\\n- **Most newer repos have low star counts** (1–4 stars), suggesting early-stage research artifacts rather than production-adopted tooling [E7](https://github.com/amazon-science/dualkv-flash-attn-for-rl)[E12](https://github.com/amazon-science/EvoMAS)[E17](https://github.com/amazon-science/adaptive-layerwise-perturbation)[E19](https://github.com/amazon-science/temporal-reasoning-dataset)[E20](https://github.com/amazon-science/PROF-GRPO)[E21](https://github.com/amazon-science/hallucination-benchmark-trivialplus)[E23](https://github.com/amazon-science/RecArena)[E24](https://github.com/amazon-science/compagent)[E28](https://github.com/amazon-science/rmir)[E34](https://github.com/amazon-science/CodeStruct)[E51](https://github.com/amazon-science/storm-referring-multi-object-grounding)[E54](https://github.com/amazon-science/acclaim)[E55](https://github.com/amazon-science/agentic-forking-path)[E60](https://github.com/amazon-science/TSFM-Biases).\\n\\n## Sources\\n\\nP1, P2, P5, P6, P7, P8, P9, P10, P11, P12, P13, P14, P15, P16, P17, P18, P19, P20, P21, P22, P23, P24, P25, P26, P27, P28, E1, E2, E3, E4, E5, E6, E7, E8, E9, E10, E11, E12, E13, E14, E15, E16, E17, E18, E19, E20, E21, E22, E23, E24, E25, E26, E27, E28, E29, E30, E31, E32, E33, E34, E35, E36, E37, E38, E39, E40, E41, E42, E43, E44, E45, E46, E47, E48, E49, E50, E51, E52, E53, E54, E55, E56, E57, E58, E59, E60, W1, W2, W3, W4\",\n  \"cites\": [\n    \"P1\", \"P2\", \"P5\", \"P6\", \"P7\", \"P8\", \"P9\", \"P10\", \"P11\", \"P12\", \"P13\", \"P14\", \"P15\", \"P16\", \"P17\", \"P18\", \"P19\", \"P20\", \"P21\", \"P22\", \"P23\", \"P24\", \"P25\", \"P26\", \"P27\", \"P28\",\n    \"E1\", \"E2\", \"E3\", \"E4\", \"E5\", \"E6\", \"E7\", \"E8\", \"E9\", \"E10\", \"E11\", \"E12\", \"E13\", \"E14\", \"E15\", \"E16\", \"E17\", \"E18\", \"E19\", \"E20\", \"E21\", \"E22\", \"E23\", \"E24\", \"E25\", \"E26\", \"E27\", \"E28\", \"E29\", \"E30\", \"E31\", \"E32\", \"E33\", \"E34\", \"E35\", \"E36\", \"E37\", \"E38\", \"E39\", \"E40\", \"E41\", \"E42\", \"E43\", \"E44\", \"E45\", \"E46\", \"E47\", \"E48\", \"E49\", \"E50\", \"E51\", \"E52\", \"E53\", \"E54\", \"E55\", \"E56\", \"E57\", \"E58\", \"E59\", \"E60\",\n    \"W1\", \"W2\", \"W3\", \"W4\"\n  ]\n}\n```","generated_at":"2026-06-10T08:03:29.623+00:00","citations":[{"url":"https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/real-world-grounding-in-agentic-ai","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/designing-ai-agents-that-know-when-to-step-back","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-agentic-ai-helps-heal-the-systems-we-cant-replace","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://github.com/amazon-science/reskill","path":null,"label":"amazon-science/reskill","type":"external"},{"url":"https://github.com/amazon-science/EvoMAS","path":null,"label":"amazon-science/EvoMAS","type":"external"},{"url":"https://github.com/amazon-science/compagent","path":null,"label":"amazon-science/compagent","type":"external"},{"url":"https://github.com/amazon-science/agentic-forking-path","path":null,"label":"amazon-science/agentic-forking-path","type":"external"},{"url":"https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents","path":null,"label":"aboutamazon.com/news","type":"external"},{"url":"https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source","path":null,"label":"labs.amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-catastrophic-is-your-llm","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-catastrophic-is-your-llm","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://huggingface.co/amazon/chronos-2","path":null,"label":"amazon/chronos-2","type":"external"},{"url":"https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/","path":null,"label":"aws.amazon.com/blogs","type":"external"},{"url":"https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://arxiv.org/pdf/2604.15261","path":null,"label":"arxiv.org/pdf","type":"external"},{"url":"https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://github.com/amazon-science/dualkv-flash-attn-for-rl","path":null,"label":"amazon-science/dualkv-flash-attn-for-rl","type":"external"},{"url":"https://github.com/amazon-science/adaptive-layerwise-perturbation","path":null,"label":"amazon-science/adaptive-layerwise-perturbation","type":"external"},{"url":"https://github.com/amazon-science/temporal-reasoning-dataset","path":null,"label":"amazon-science/temporal-reasoning-dataset","type":"external"},{"url":"https://github.com/amazon-science/PROF-GRPO","path":null,"label":"amazon-science/PROF-GRPO","type":"external"},{"url":"https://github.com/amazon-science/hallucination-benchmark-trivialplus","path":null,"label":"amazon-science/hallucination-benchmark-trivialplus","type":"external"},{"url":"https://github.com/amazon-science/RecArena","path":null,"label":"amazon-science/RecArena","type":"external"},{"url":"https://github.com/amazon-science/SWAN","path":null,"label":"amazon-science/SWAN","type":"external"},{"url":"https://github.com/amazon-science/rmir","path":null,"label":"amazon-science/rmir","type":"external"},{"url":"https://github.com/amazon-science/expert-upcycling","path":null,"label":"amazon-science/expert-upcycling","type":"external"},{"url":"https://github.com/amazon-science/CodeStruct","path":null,"label":"amazon-science/CodeStruct","type":"external"},{"url":"https://github.com/amazon-science/TransitionFlowMatching","path":null,"label":"amazon-science/TransitionFlowMatching","type":"external"},{"url":"https://github.com/amazon-science/storm-referring-multi-object-grounding","path":null,"label":"amazon-science/storm-referring-multi-object-grounding","type":"external"},{"url":"https://github.com/amazon-science/acclaim","path":null,"label":"amazon-science/acclaim","type":"external"},{"url":"https://github.com/amazon-science/papercode-coordinating-spot-and-contracts","path":null,"label":"amazon-science/papercode-coordinating-spot-and-contracts","type":"external"},{"url":"https://github.com/amazon-science/TSFM-Biases","path":null,"label":"amazon-science/TSFM-Biases","type":"external"},{"url":"https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context","path":null,"label":"amazon/gpt-oss-20b-p-eagle-long-context","type":"external"},{"url":"https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context","path":null,"label":"amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context","type":"external"},{"url":"https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context","path":null,"label":"amazon/gpt-oss-120b-p-eagle-long-context","type":"external"},{"url":"https://huggingface.co/amazon/gpt-oss-120b-p-eagle","path":null,"label":"amazon/gpt-oss-120b-p-eagle","type":"external"},{"url":"https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct","path":null,"label":"amazon/GKA-primed-HQwen3-32B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct","path":null,"label":"amazon/Mamba2-primed-HQwen3-8B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner","path":null,"label":"amazon/GKA-primed-HQwen3-8B-Reasoner","type":"external"},{"url":"https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner","path":null,"label":"amazon/GKA-primed-HQwen3-32B-Reasoner","type":"external"},{"url":"https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct","path":null,"label":"amazon/GDN-primed-HQwen3-32B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct","path":null,"label":"amazon/GKA-primed-HQwen3-8B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct","path":null,"label":"amazon/GDN-primed-HQwen3-8B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct","path":null,"label":"amazon/BMOJOF-primed-HQwen3-8B-Instruct","type":"external"},{"url":"https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner","path":null,"label":"amazon/GDN-primed-HQwen3-8B-Reasoner","type":"external"},{"url":"https://github.com/amazon-science/concurry/releases/tag/v0.13.2","path":null,"label":"amazon-science/concurry","type":"external"},{"url":"https://github.com/amazon-science/concurry/releases/tag/v0.13.1","path":null,"label":"amazon-science/concurry","type":"external"},{"url":"https://github.com/amazon-science/azcausal/releases/tag/v0.2.5","path":null,"label":"amazon-science/azcausal","type":"external"},{"url":"https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3","path":null,"label":"amazon-science/azcausal","type":"external"},{"url":"https://github.com/amazon-science/uniqsketch/releases/tag/v1.2.1","path":null,"label":"amazon-science/uniqsketch","type":"external"},{"url":"https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/real-world-grounding-in-agentic-ai","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/building-trust-into-ai","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/building-trust-into-ai","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/formally-verified-aes-xts-the-first-aes-algorithm-to-join-s2n-bignum","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/news/how-academic-collaboration-delivers-real-world-security-to-amazon-customers","path":null,"label":"amazon.science/news","type":"external"},{"url":"https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/optimizing-lora-target-module-selection-for-efficient-fine-tuning","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/why-a-12-year-old-forecasting-paper-has-stood-the-test-of-time","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design","path":null,"label":"amazon.science/news","type":"external"},{"url":"https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design","path":null,"label":"amazon.science/news","type":"external"},{"url":"https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/how-ai-is-changing-the-nature-of-mathematical-research","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/ground-truth-is-a-process-not-a-dataset","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://aws.amazon.com/blogs/machine-learning/evaluate-your-amazon-nova-sonic-voice-agent-at-scale-no-microphone-required/","path":null,"label":"aws.amazon.com/blogs","type":"external"},{"url":"https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced","path":null,"label":"amazon.science/research-awards","type":"external"},{"url":"https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced","path":null,"label":"amazon.science/research-awards","type":"external"}],"provenance":{"provider":"deepseek","model":"deepseek-v4-pro","workflow":"onlylabs-deepagents-analysis-v3","agent":"deepagents"},"evidence":{"total":92,"pages":28,"events":140,"web":4,"signal_desks":{"forks":0,"repos":19,"hiring":0,"talking":22,"releases":19},"data_radar_lanes":{"data":10,"evals":6,"safety":5,"product":2,"infrastructure":8},"data_radar_matches":21}},{"org_slug":"anthropic","url":"https://onlylabs.fyi/analysis/anthropic","json_url":"https://onlylabs.fyi/analysis/anthropic/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/anthropic/evidence.json","dossier_url":"https://onlylabs.fyi/labs/anthropic","org":{"slug":"anthropic","name":"Anthropic","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.anthropic.com"},"title":"Anthropic analysis","summary":"Anthropic in mid-2026 is executing a multi-front strategy: securing unprecedented compute capacity across AWS, SpaceX, Google, and Microsoft to train and serve models at scale; building out a mature enterprise GTM organization with regional hubs, vertical specialists, and a dedicated AI services company; shipping a rapidly iterating agent platform spanning Claude Code, Agent SDKs, and Managed Agents; restricting…","markdown":"```json\n{\n  \"content\": \"## Thesis\\nAnthropic in mid-2026 is executing a multi-front strategy: securing unprecedented compute capacity across AWS, SpaceX, Google, and Microsoft to train and serve models at scale [P14](https://www.anthropic.com/news/anthropic-amazon-compute)[P25](https://www.anthropic.com/news/higher-limits-spacex); building out a mature enterprise GTM organization with regional hubs, vertical specialists, and a dedicated AI services company [P15](https://www.anthropic.com/news/anthropic-nec)[P17](https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand)[P22](https://www.anthropic.com/news/enterprise-ai-services-company)[E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008)[E17](https://job-boards.greenhouse.io/anthropic/jobs/5222289008); shipping a rapidly iterating agent platform spanning Claude Code, Agent SDKs, and Managed Agents [P19](https://www.anthropic.com/engineering/managed-agents)[E6](https://github.com/anthropics/claude-code/releases/tag/v2.1.168)[E5](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168)[E18](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93); restricting their most capable model (Mythos Preview) while releasing Opus 4.7 with differential cyber safeguards as a safer stepping stone [P26](https://www.anthropic.com/news/claude-opus-4-7)[P27](https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf)[P7](https://www.anthropic.com/research/glasswing-initial-update); and investing heavily in alignment tooling, interpretability research, and safety data infrastructure [P2](https://www.anthropic.com/research/donating-open-source-petri)[P6](https://www.anthropic.com/research/teaching-claude-why)[P10](https://www.anthropic.com/research/natural-language-autoencoders)[P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers)[E47](https://job-boards.greenhouse.io/anthropic/jobs/5247156008).\\n\\n## Signal desks\\n\\n### Hiring\\n- **Enterprise GTM buildout**: Enterprise Account Executive for Manufacturing in London [E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008), Manager of Solutions Architecture for Applied AI (Enterprise Tech) in SF/NYC [E17](https://job-boards.greenhouse.io/anthropic/jobs/5222289008), Applied AI Architect Lead for EMEA Commercial in Dublin [E57](https://job-boards.greenhouse.io/anthropic/jobs/5201715008), and Startup Partnerships Lead in SF/NYC [E27](https://job-boards.greenhouse.io/anthropic/jobs/5235692008) signal a broad push to commercialize Claude across verticals and geographies. Senior Manager for Order Management – Partnership NPI & Automation [E28](https://job-boards.greenhouse.io/anthropic/jobs/5205738008) and Commercial Counsel for Platform & Marketplace [E33](https://job-boards.greenhouse.io/anthropic/jobs/5208289008) reinforce partnership and marketplace scaling.\\n- **Platform infrastructure roles**: Product Manager for Compute Platform in SF/NYC/Seattle [E32](https://job-boards.greenhouse.io/anthropic/jobs/5124623008), Engineering Manager for Safeguards Data Infrastructure in NYC [E47](https://job-boards.greenhouse.io/anthropic/jobs/5247156008), and Engineering Manager for GRC Platform in SF/NYC/Seattle [E16](https://job-boards.greenhouse.io/anthropic/jobs/4980335008) indicate dedicated buildout of the underlying compute, data, and compliance infrastructure layers.\\n- **Public sector and government**: Staff+ Software Engineer for Public Sector (Remote/DC) [E45](https://job-boards.greenhouse.io/anthropic/jobs/5205704008) and Lead for Government Incentives & Economic Development in SF [E29](https://job-boards.greenhouse.io/anthropic/jobs/5205505008) point to a formal government vertical with economic development and incentive strategy.\\n- **Product and commercial operations**: Product Manager for Claude Code Model Performance in SF/Seattle [E46](https://job-boards.greenhouse.io/anthropic/jobs/5247640008) suggests dedicated product ownership for agent coding quality, directly relevant to the issues surfaced in the April 2026 postmortem [P20](https://www.anthropic.com/engineering/april-23-postmortem). Strategy & Operations Lead for Marketing [E30](https://job-boards.greenhouse.io/anthropic/jobs/5248946008) and Financial Reporting Accountant [P9](https://job-boards.greenhouse.io/anthropic/jobs/5248494008) reflect scaling of commercial operations and financial reporting maturity.\\n\\n### Forks\\n- No cited evidence in this pack.\\n\\n### Releases\\n- **Claude Code and agent platform**: Claude Code shipped multiple releases in this pack (v2.1.166–v2.1.168) [E6](https://github.com/anthropics/claude-code/releases/tag/v2.1.168)[E20](https://github.com/anthropics/claude-code/releases/tag/v2.1.167)[E25](https://github.com/anthropics/claude-code/releases/tag/v2.1.166), alongside coordinated Agent SDK releases for Python (v0.2.92–v0.2.93) [E18](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93)[E22](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.92) and TypeScript (v0.3.166–v0.3.168) [E5](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168)[E21](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.167)[E24](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.166), plus the Claude Code GitHub Action for CI/CD integration (v1.0.138–v1.0.140) [E4](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140)[E19](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.139)[E23](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.138). This cadence reflects a rapid agent-product iteration cycle.\\n- **Multi-platform API SDKs**: SDKs ship across eight surfaces — Python [E2](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.1)[E13](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.0)[E31](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.106.0), TypeScript core plus AWS Bedrock, Vertex AI, and Microsoft Foundry variants [E12](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1)[E14](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1)[E15](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.102.0)[E36](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0)[E37](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0)[E38](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0)[E39](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0)[E40](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.101.0), Java [E10](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.38.0)[E35](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.37.0)[E49](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.36.0), Go [E11](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.48.0)[E44](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.47.0), Ruby [E7](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.46.0)[E43](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.45.0), C# [E8](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.27.0)[E42](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.26.0)[E48](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.25.0), and PHP [E9](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.27.0)[E41](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.26.0) — on a near-daily cadence, signaling a developer-platform strategy targeting all major enterprise ecosystems.\\n- **Internal tooling**: buffa v0.5.0 [P28](https://github.com/anthropics/buffa/releases/tag/v0.5.0) released with Rust protobuf codegen improvements including BSR Cargo-SDK compatibility and zero-copy decode, suggesting internal gRPC/protobuf service architecture investment.\\n\\n### Talking\\n- **Cybersecurity as a flagship differentiator**: The disruption of the first reported AI-orchestrated cyber espionage campaign attributed to Chinese state-sponsored group GTG-1002 [P5](https://assets.anthropic.com/m/ec212e6566a0d47/original/Disrupting-the-first-reported-AI-orchestrated-cyber-espionage-campaign.pdf) and Project Glasswing's finding of 10,000+ high- or critical-severity vulnerabilities across systemically important software with ~50 partners [P7](https://www.anthropic.com/research/glasswing-initial-update) position Anthropic as uniquely capable in AI-driven cyber offense and defense. The Mythos Preview System Card devotes significant space to cybersecurity evaluations and the release decision to restrict access [P27](https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf).\\n- **Geopolitical positioning**: The 2028 AI Leadership paper frames US-China competition around compute advantage, export controls, and Chinese distillation attacks, presenting two scenarios hinging on whether policymakers tighten loopholes [P4](https://www.anthropic.com/research/2028-ai-leadership). The election safeguards update emphasizing political neutrality evaluation methodology (Opus 4.7 at 95%, Sonnet 4.6 at 96%) [P21](https://www.anthropic.com/news/election-safeguards-update) complements the geopolitical narrative with domestic democratic resilience framing.\\n- **Alignment and interpretability breakthroughs**: Teaching Claude \\\"why\\\" through principled OOD alignment training — agentic misalignment eliminated entirely since Claude Haiku 4.5, down from Opus 4's 96% blackmail rate [P6](https://www.anthropic.com/research/teaching-claude-why). Natural Language Autoencoders (NLAs) convert model activations into readable text, applied during Opus 4.6 and Mythos Preview safety testing [P10](https://www.anthropic.com/research/natural-language-autoencoders). The persona selection model theory argues human-like behavior is a default outcome of pretraining [P11](https://www.anthropic.com/research/persona-selection-model). Next-generation Constitutional Classifiers reduce compute overhead while maintaining jailbreak defense [P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers).\\n- **Enterprise product narrative**: Managed Agents as a hosted service with stable interfaces decoupling the \\\"brain from the hands\\\" for long-horizon agent work [P19](https://www.anthropic.com/engineering/managed-agents), the April 23 Claude Code quality postmortem transparently detailing three separate regressions and their remediation [P20](https://www.anthropic.com/engineering/april-23-postmortem), creative tool connectors spanning Ableton, Adobe Creative Cloud, Blender, Autodesk Fusion, and more [P24](https://www.anthropic.com/news/claude-for-creative-work), and the NEC partnership to build Japan's largest AI-native engineering organization (~30,000 employees) [P15](https://www.anthropic.com/news/anthropic-nec) all reinforce a maturing enterprise product story.\\n- **Economic and societal research**: The Anthropic Institute (TAI) agenda covers economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D, with commitments to share more granular Economic Index data and internal evidence of AI-accelerated work [P3](https://www.anthropic.com/research/anthropic-institute-agenda). The coding agents in social sciences survey (n=1,260) finds 81% have tried AI chatbots but only 20% use coding agents, with sharp gender and institutional disparities [P8](https://www.anthropic.com/research/coding-agents-social-sciences).\\n- **Additional research themes signaled by post titles**: Agent autonomy measurement [E50](https://www.anthropic.com/research/measuring-agent-autonomy), values alignment in the wild [E51](https://www.anthropic.com/research/values-wild), emergent misalignment and reward hacking [E58](https://www.anthropic.com/research/emergent-misalignment-reward-hacking), how AI is transforming work at Anthropic [E52](https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic), assistant personality/behavior [E53](https://www.anthropic.com/research/assistant-axis), introspection [E54](https://www.anthropic.com/research/introspection), productivity estimation [E55](https://www.anthropic.com/research/estimating-productivity-gains), emotion concepts [E59](https://www.anthropic.com/research/emotion-concepts-function), personal guidance features [E56](https://www.anthropic.com/research/claude-personal-guidance), containment engineering [E26](https://www.anthropic.com/engineering/how-we-contain-claude), and agents in biology [E1](https://www.anthropic.com/research/agents-in-biology) round out a broad research program across safety, product, and scientific applications.\\n\\n## Shipping\\nAnthropic's shipping velocity is concentrated in three areas:\\n\\n1. **Claude Code and agent platform**: Claude Code itself on a rapid release cadence (v2.1.166–v2.1.168 tracked in this pack) [E6](https://github.com/anthropics/claude-code/releases/tag/v2.1.168)[E20](https://github.com/anthropics/claude-code/releases/tag/v2.1.167)[E25](https://github.com/anthropics/claude-code/releases/tag/v2.1.166) alongside coordinated Agent SDK releases for Python [E18](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93)[E22](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.92) and TypeScript [E5](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168)[E21](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.167)[E24](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.166), plus the Claude Code GitHub Action [E4](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140)[E19](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.139)[E23](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.138). Managed Agents launched as a hosted service abstracting harness complexity behind stable interfaces [P19](https://www.anthropic.com/engineering/managed-agents).\\n\\n2. **Multi-platform SDK ecosystem**: SDKs ship across eight language and platform surfaces (Python, TypeScript core + AWS Bedrock + Vertex AI + Microsoft Foundry variants, Java, Go, Ruby, C#, PHP) on a near-daily cadence [E2](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.1)[E7](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.46.0)[E8](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.27.0)[E9](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.27.0)[E10](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.38.0)[E11](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.48.0)[E12](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1)[E13](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.0)[E14](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1)[E15](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.102.0)[E31](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.106.0)[E35](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.37.0)[E36](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0)[E37](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0)[E38](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0)[E39](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0)[E40](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.101.0)[E41](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.26.0)[E42](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.26.0)[E43](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.45.0)[E44](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.47.0)[E48](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.25.0)[E49](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.36.0).\\n\\n3. **Model releases**: Claude Opus 4.7 shipped April 2026 with improved software engineering, better vision capabilities, and differential cyber safeguards — the first model released with learnings from Mythos-class safeguards [P26](https://www.anthropic.com/news/claude-opus-4-7). Claude Mythos Preview shipped as a restricted-access model for defensive cybersecurity partners only [P27](https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf)[P7](https://www.anthropic.com/research/glasswing-initial-update).\\n\\n## Research themes\\nAnthropic's public research clusters around five themes:\\n\\n- **Interpretability**: NLAs convert model activations into natural-language text, revealing internal reasoning during safety testing for Opus 4.6 and Mythos Preview [P10](https://www.anthropic.com/research/natural-language-autoencoders).\\n- **Alignment training**: Agentic misalignment eliminated through principled OOD training (constitutions, fictional stories) rather than eval-distribution overfitting; training on demonstrations alone proved insufficient — teaching Claude to explain *why* was more effective [P6](https://www.anthropic.com/research/teaching-claude-why). Constitutional Classifiers evolved to a second generation with reduced compute overhead and improved refusal rates [P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers).\\n- **Persona and behavior**: The persona selection model theory argues human-like behavior is a default outcome of pretraining, not purely instilled through character training [P11](https://www.anthropic.com/research/persona-selection-model).\\n- **Scientific capabilities**: BioMysteryBench for evaluating bioinformatics research capabilities [P1](https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench); agents in biology [E1](https://www.anthropic.com/research/agents-in-biology); making Claude a chemist [E34](https://www.anthropic.com/research/making-claude-a-chemist).\\n- **Agent eval methodology**: Infrastructure configuration alone can swing agentic coding benchmark scores by 6 percentage points (p < 0.01) on Terminal-Bench 2.0 — more than the leaderboard gap between top models [P13](https://www.anthropic.com/engineering/infrastructure-noise).\\n\\n## Hiring and scaling\\nAnthropic's hiring signals a transition from research lab to multi-geography enterprise platform company:\\n\\n- **Geographic expansion**: Roles in London [E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008), Dublin [E57](https://job-boards.greenhouse.io/anthropic/jobs/5201715008), and a Sydney office opening with a dedicated GM for Australia and New Zealand [P17](https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand) complement the core SF/NYC/Seattle hubs and DC presence [E45](https://job-boards.greenhouse.io/anthropic/jobs/5205704008)[E33](https://job-boards.greenhouse.io/anthropic/jobs/5208289008). Compute infrastructure expansion includes inference capacity in Asia and Europe [P14](https://www.anthropic.com/news/anthropic-amazon-compute).\\n- **Enterprise GTM**: Solutions architecture management [E17](https://job-boards.greenhouse.io/anthropic/jobs/5222289008), applied AI architects for EMEA [E57](https://job-boards.greenhouse.io/anthropic/jobs/5201715008), startup partnerships [E27](https://job-boards.greenhouse.io/anthropic/jobs/5235692008), manufacturing vertical AE [E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008), and partnership operations/NPI [E28](https://job-boards.greenhouse.io/anthropic/jobs/5205738008) indicate a maturing enterprise sales motion with vertical specialization.\\n- **Platform infrastructure**: Compute Platform PM [E32](https://job-boards.greenhouse.io/anthropic/jobs/5124623008), Safeguards Data Infrastructure engineering manager [E47](https://job-boards.greenhouse.io/anthropic/jobs/5247156008), GRC Platform engineering manager [E16](https://job-boards.greenhouse.io/anthropic/jobs/4980335008), and Claude Code model performance PM [E46](https://job-boards.greenhouse.io/anthropic/jobs/5247640008) show dedicated investment in the infrastructure and product layers beneath the models.\\n- **Government and public sector**: A dedicated public sector software engineering role [E45](https://job-boards.greenhouse.io/anthropic/jobs/5205704008) and government incentives lead [E29](https://job-boards.greenhouse.io/anthropic/jobs/5205505008) suggest a formal government vertical strategy.\\n- **Financial reporting**: The Financial Reporting Accountant role with Workiva and US GAAP expertise [P9](https://job-boards.greenhouse.io/anthropic/jobs/5248494008) could signal preparation for public company financial reporting, though not explicitly stated.\\n\\n## Data-business implications\\nBased on cited evidence only:\\n\\n- **Evals and benchmarking**: Anthropic's finding that infrastructure noise can swing agentic coding evals by 6pp [P13](https://www.anthropic.com/engineering/infrastructure-noise) creates demand for standardized, reproducible eval infrastructure and runtime enforcement. The BioMysteryBench [P1](https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench) and broader scientific eval work suggest growing need for domain-specific evaluation datasets and tooling, particularly in bioinformatics.\\n- **Data infrastructure**: The Safeguards Data Infrastructure engineering manager role [E47](https://job-boards.greenhouse.io/anthropic/jobs/5247156008) and the Constitutional Classifiers' reliance on synthetic data generated from constitutions [P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers) indicate internal investment in safety-specific data pipelines. The Managed Agents architecture — virtualizing sessions as append-only logs [P19](https://www.anthropic.com/engineering/managed-agents) — implies significant logging, storage, and data retention infrastructure requirements for agent observability.\\n- **Deployment and multi-cloud**: Anthropic's SDK coverage across AWS Bedrock, Vertex AI, and Microsoft Foundry [E12](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1)[E14](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1)[E36](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0)[E37](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0)[E38](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0)[E39](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0) plus compute deals spanning AWS Trainium, Google TPUs, and NVIDIA GPUs [P14](https://www.anthropic.com/news/anthropic-amazon-compute)[P25](https://www.anthropic.com/news/higher-limits-spacex) signals a multi-cloud, multi-silicon deployment model. The Claude Platform on AWS with same-account, same-controls integration [P14](https://www.anthropic.com/news/anthropic-amazon-compute) and the NEC partnership for in-country deployment in Japan [P15](https://www.anthropic.com/news/anthropic-nec) suggest enterprise demand for managed, region-specific, and sovereign deployment options.\\n- **Tooling and agent infrastructure**: Claude Code, Agent SDKs, and Managed Agents [P19](https://www.anthropic.com/engineering/managed-agents)[E6](https://github.com/anthropics/claude-code/releases/tag/v2.1.168)[E5](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168)[E18](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93) represent a growing surface area for developer tooling, CI/CD integration (GitHub Action) [E4](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140), and agent monitoring/observability. The postmortem on Claude Code quality issues — tracing degradation to three separate changes including a default reasoning effort reduction and a context-clearing bug [P20](https://www.anthropic.com/engineering/april-23-postmortem) — highlights the operational complexity of agent products and the need for testing and quality observability infrastructure.\\n- **Safety and alignment products**: Petri 3.0 (open-source alignment testing, now with the \\\"Dish\\\" add-on for realism and Bloom integration for deeper assessments) [P2](https://www.anthropic.com/research/donating-open-source-petri), Constitutional Classifiers [P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers), and the Cyber Verification Program for Opus 4.7 [P26](https://www.anthropic.com/news/claude-opus-4-7) suggest Anthropic is productizing safety infrastructure for external use, not just treating it as internal research.\\n- **Product and GTM**: The enterprise AI services company with Blackstone, H&F, and Goldman Sachs, backed by GIC and Sequoia, extends delivery capacity for mid-market enterprises that lack in-house AI resources [P22](https://www.anthropic.com/news/enterprise-ai-services-company). Creative tool connectors spanning Ableton, Adobe, Blender, Autodesk Fusion, SketchUp, and Splice open a new vertical (creative professionals) [P24](https://www.anthropic.com/news/claude-for-creative-work). The NEC partnership targets finance, manufacturing, and local government in Japan [P15](https://www.anthropic.com/news/anthropic-nec). These all create demand for vertical-specific product customization, deployment, and support.\\n- **Compute infrastructure**: The $100B+ AWS commitment spanning Graviton and Trainium2–Trainium4, nearly 1GW of Trainium2/Trainium3 capacity by end of 2026 [P14](https://www.anthropic.com/news/anthropic-amazon-compute); SpaceX Colossus 1 (300MW, 220K+ NVIDIA GPUs) available within a month of May 2026 [P25](https://www.anthropic.com/news/higher-limits-spacex); Google/Broadcom 5GW beginning in 2027; Microsoft $30B Azure; and $50B Fluidstack investment [P25](https://www.anthropic.com/news/higher-limits-spacex) represent one of the largest infrastructure buildouts in AI, with implications for silicon diversity, data center supply chains, and networking.\\n\\n## Traction highlights\\n- **Enterprise adoption**: 100,000+ customers running Claude on Amazon Bedrock [P14](https://www.anthropic.com/news/anthropic-amazon-compute). NEC deploying Claude to ~30,000 employees as Japan's largest AI-native engineering organization [P15](https://www.anthropic.com/news/anthropic-nec).\\n- **Cybersecurity impact**: Project Glasswing with ~50 partners found 10,000+ high- or critical-severity vulnerabilities in systemically important software within its first weeks [P7](https://www.anthropic.com/research/glasswing-initial-update).\\n- **Compute scale**: Over 1 million Trainium2 chips currently used for training and serving Claude [P14](https://www.anthropic.com/news/anthropic-amazon-compute). Access to 300MW+ (220K+ NVIDIA GPUs) at SpaceX Colossus 1 within a month [P25](https://www.anthropic.com/news/higher-limits-spacex). Nearly 1GW of Trainium2/Trainium3 capacity coming online by end of 2026, with Trainium4 optionally available [P14](https://www.anthropic.com/news/anthropic-amazon-compute).\\n- **Model safety performance**: Agentic misalignment eliminated (0% blackmail rate) since Claude Haiku 4.5, compared to Opus 4's 96% rate [P6](https://www.anthropic.com/research/teaching-claude-why). Opus 4.7 scored 95% and Sonnet 4.6 scored 96% on political bias/impartiality evaluation [P21](https://www.anthropic.com/news/election-safeguards-update).\\n- **International expansion**: Sydney office officially opened with dedicated GM [P17](https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand). EMEA commercial presence building via Dublin [E57](https://job-boards.greenhouse.io/anthropic/jobs/5201715008) and London [E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008). Inference expansion in Asia and Europe [P14](https://www.anthropic.com/news/anthropic-amazon-compute).\\n- **Alignment tooling adoption**: Petri used externally by the UK AISI for model evaluation [P2](https://www.anthropic.com/research/donating-open-source-petri). NLAs released as open-source code with a Neuronpedia collaboration for interactive exploration [P10](https://www.anthropic.com/research/natural-language-autoencoders). Constitutional Classifiers detailed in a published paper [P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers).\\n\\n## Sources\\n[P1](https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench) Evaluating Claude For Bioinformatics With BioMysteryBench (Apr 29, 2026) · https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench\\n[P2](https://www.anthropic.com/research/donating-open-source-petri) Donating Open Source Petri (May 7, 2026) · https://www.anthropic.com/research/donating-open-source-petri\\n[P3](https://www.anthropic.com/research/anthropic-institute-agenda) Anthropic Institute Agenda (May 7, 2026) · https://www.anthropic.com/research/anthropic-institute-agenda\\n[P4](https://www.anthropic.com/research/2028-ai-leadership) 2028: Two scenarios for global AI leadership (May 14, 2026) · https://www.anthropic.com/research/2028-ai-leadership\\n[P5](https://assets.anthropic.com/m/ec212e6566a0d47/original/Disrupting-the-first-reported-AI-orchestrated-cyber-espionage-campaign.pdf) Disrupting the first reported AI-orchestrated cyber espionage campaign (Nov 2025) · https://assets.anthropic.com/m/ec212e6566a0d47/original/Disrupting-the-first-reported-AI-orchestrated-cyber-espionage-campaign.pdf\\n[P6](https://www.anthropic.com/research/teaching-claude-why) Teaching Claude why (May 8, 2026) · https://www.anthropic.com/research/teaching-claude-why\\n[P7](https://www.anthropic.com/research/glasswing-initial-update) Project Glasswing: An initial update (May 22, 2026) · https://www.anthropic.com/research/glasswing-initial-update\\n[P8](https://www.anthropic.com/research/coding-agents-social-sciences) Coding agents in the social sciences (May 27, 2026) · https://www.anthropic.com/research/coding-agents-social-sciences\\n[P9](https://job-boards.greenhouse.io/anthropic/jobs/5248494008) Financial Reporting Accountant job · https://job-boards.greenhouse.io/anthropic/jobs/5248494008\\n[P10](https://www.anthropic.com/research/natural-language-autoencoders) Natural Language Autoencoders (May 7, 2026) · https://www.anthropic.com/research/natural-language-autoencoders\\n[P11](https://www.anthropic.com/research/persona-selection-model) The persona selection model (Feb 23, 2026) · https://www.anthropic.com/research/persona-selection-model\\n[P12](https://www.anthropic.com/research/next-generation-constitutional-classifiers) Next-generation Constitutional Classifiers (Jan 9, 2026) · https://www.anthropic.com/research/next-generation-constitutional-classifiers\\n[P13](https://www.anthropic.com/engineering/infrastructure-noise) Quantifying infrastructure noise in agentic coding evals (Feb 5, 2026) · https://www.anthropic.com/engineering/infrastructure-noise\\n[P14](https://www.anthropic.com/news/anthropic-amazon-compute) Anthropic and Amazon expand collaboration for up to 5GW of new compute (Apr 20, 2026) · https://www.anthropic.com/news/anthropic-amazon-compute\\n[P15](https://www.anthropic.com/news/anthropic-nec) Anthropic and NEC partner to build AI-native engineering at scale in Japan (Apr 24, 2026) · https://www.anthropic.com/news/anthropic-nec\\n[P17](https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand) Anthropic names Theo Hourmouzis GM of Australia & New Zealand (Apr 27, 2026) · https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand\\n[P18](https://www.anthropic.com/news/the-long-term-benefit-trust) The Long-Term Benefit Trust (Sep 19, 2023) · https://www.anthropic.com/news/the-long-term-benefit-trust\\n[P19](https://www.anthropic.com/engineering/managed-agents) Scaling Managed Agents: Decoupling the brain from the hands (Apr 8, 2026) · https://www.anthropic.com/engineering/managed-agents\\n[P20](https://www.anthropic.com/engineering/april-23-postmortem) An update on recent Claude Code quality reports (Apr 23, 2026) · https://www.anthropic.com/engineering/april-23-postmortem\\n[P21](https://www.anthropic.com/news/election-safeguards-update) An update on our election safeguards (Apr 24, 2026) · https://www.anthropic.com/news/election-safeguards-update\\n[P22](https://www.anthropic.com/news/enterprise-ai-services-company) Building a new enterprise AI services company with Blackstone, H&F, and Goldman Sachs (May 4, 2026) · https://www.anthropic.com/news/enterprise-ai-services-company\\n[P24](https://www.anthropic.com/news/claude-for-creative-work) Claude for Creative Work (Apr 28, 2026) · https://www.anthropic.com/news/claude-for-creative-work\\n[P25](https://www.anthropic.com/news/higher-limits-spacex) Higher usage limits for Claude and a compute deal with SpaceX (May 6, 2026) · https://www.anthropic.com/news/higher-limits-spacex\\n[P26](https://www.anthropic.com/news/claude-opus-4-7) Introducing Claude Opus 4.7 (Apr 16, 2026) · https://www.anthropic.com/news/claude-opus-4-7\\n[P27](https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf) System Card: Claude Mythos Preview (Apr 7, 2026) · https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf\\n[P28](https://github.com/anthropics/buffa/releases/tag/v0.5.0) anthropics/buffa v0.5.0 (May 5, 2026) · https://github.com/anthropics/buffa/releases/tag/v0.5.0\\n[E1](https://www.anthropic.com/research/agents-in-biology) Agents In Biology post · https://www.anthropic.com/research/agents-in-biology\\n[E3](https://job-boards.greenhouse.io/anthropic/jobs/5237973008) Enterprise Account Executive, Manufacturing job · https://job-boards.greenhouse.io/anthropic/jobs/5237973008\\n[E4](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140) anthropics/claude-code-action v1.0.140 · https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140\\n[E5](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168) anthropics/claude-agent-sdk-typescript v0.3.168 · https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168\\n[E6](https://github.com/anthropics/claude-code/releases/tag/v2.1.168) anthropics/claude-code v2.1.168 · https://github.com/anthropics/claude-code/releases/tag/v2.1.168\\n[E7](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.46.0) anthropics/anthropic-sdk-ruby v1.46.0 · https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.46.0\\n[E8](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.27.0) anthropics/anthropic-sdk-csharp Anthropic-v12.27.0 · https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.27.0\\n[E9](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.27.0) anthropics/anthropic-sdk-php v0.27.0 · https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.27.0\\n[E10](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.38.0) anthropics/anthropic-sdk-java v2.38.0 · https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.38.0\\n[E11](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.48.0) anthropics/anthropic-sdk-go v1.48.0 · https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.48.0\\n[E12](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1) anthropics/anthropic-sdk-typescript aws-sdk-v0.4.1 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1\\n[E13](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.0) anthropics/anthropic-sdk-python v0.107.0 · https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.0\\n[E14](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1) anthropics/anthropic-sdk-typescript bedrock-sdk-v0.30.1 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1\\n[E15](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.102.0) anthropics/anthropic-sdk-typescript sdk-v0.102.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.102.0\\n[E16](https://job-boards.greenhouse.io/anthropic/jobs/4980335008) Engineering Manager, GRC Platform job · https://job-boards.greenhouse.io/anthropic/jobs/4980335008\\n[E17](https://job-boards.greenhouse.io/anthropic/jobs/5222289008) Manager of Solutions Architecture, Applied AI (Enterprise Tech) job · https://job-boards.greenhouse.io/anthropic/jobs/5222289008\\n[E18](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93) anthropics/claude-agent-sdk-python v0.2.93 · https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93\\n[E19](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.139) anthropics/claude-code-action v1.0.139 · https://github.com/anthropics/claude-code-action/releases/tag/v1.0.139\\n[E20](https://github.com/anthropics/claude-code/releases/tag/v2.1.167) anthropics/claude-code v2.1.167 · https://github.com/anthropics/claude-code/releases/tag/v2.1.167\\n[E21](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.167) anthropics/claude-agent-sdk-typescript v0.3.167 · https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.167\\n[E22](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.92) anthropics/claude-agent-sdk-python v0.2.92 · https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.92\\n[E23](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.138) anthropics/claude-code-action v1.0.138 · https://github.com/anthropics/claude-code-action/releases/tag/v1.0.138\\n[E24](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.166) anthropics/claude-agent-sdk-typescript v0.3.166 · https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.166\\n[E25](https://github.com/anthropics/claude-code/releases/tag/v2.1.166) anthropics/claude-code v2.1.166 · https://github.com/anthropics/claude-code/releases/tag/v2.1.166\\n[E26](https://www.anthropic.com/engineering/how-we-contain-claude) How We Contain Claude post · https://www.anthropic.com/engineering/how-we-contain-claude\\n[E27](https://job-boards.greenhouse.io/anthropic/jobs/5235692008) Startup Partnerships Lead job · https://job-boards.greenhouse.io/anthropic/jobs/5235692008\\n[E28](https://job-boards.greenhouse.io/anthropic/jobs/5205738008) Senior Manager, Order Management – Partnership NPI & Automation job · https://job-boards.greenhouse.io/anthropic/jobs/5205738008\\n[E29](https://job-boards.greenhouse.io/anthropic/jobs/5205505008) Lead, Government Incentives & Economic Development job · https://job-boards.greenhouse.io/anthropic/jobs/5205505008\\n[E30](https://job-boards.greenhouse.io/anthropic/jobs/5248946008) Strategy & Operations Lead, Marketing job · https://job-boards.greenhouse.io/anthropic/jobs/5248946008\\n[E31](https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.106.0) anthropics/anthropic-sdk-python v0.106.0 · https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.106.0\\n[E32](https://job-boards.greenhouse.io/anthropic/jobs/5124623008) Product Manager, Compute Platform job · https://job-boards.greenhouse.io/anthropic/jobs/5124623008\\n[E33](https://job-boards.greenhouse.io/anthropic/jobs/5208289008) Commercial Counsel, Platform & Marketplace job · https://job-boards.greenhouse.io/anthropic/jobs/5208289008\\n[E34](https://www.anthropic.com/research/making-claude-a-chemist) Making Claude A Chemist post · https://www.anthropic.com/research/making-claude-a-chemist\\n[E35](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.37.0) anthropics/anthropic-sdk-java v2.37.0 · https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.37.0\\n[E36](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0) anthropics/anthropic-sdk-typescript aws-sdk-v0.4.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0\\n[E37](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0) anthropics/anthropic-sdk-typescript foundry-sdk-v0.3.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0\\n[E38](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0) anthropics/anthropic-sdk-typescript bedrock-sdk-v0.30.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0\\n[E39](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0) anthropics/anthropic-sdk-typescript vertex-sdk-v0.17.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0\\n[E40](https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.101.0) anthropics/anthropic-sdk-typescript sdk-v0.101.0 · https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.101.0\\n[E41](https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.26.0) anthropics/anthropic-sdk-php v0.26.0 · https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.26.0\\n[E42](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.26.0) anthropics/anthropic-sdk-csharp Anthropic-v12.26.0 · https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.26.0\\n[E43](https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.45.0) anthropics/anthropic-sdk-ruby v1.45.0 · https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.45.0\\n[E44](https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.47.0) anthropics/anthropic-sdk-go v1.47.0 · https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.47.0\\n[E45](https://job-boards.greenhouse.io/anthropic/jobs/5205704008) Staff+ Software Engineer, Public Sector job · https://job-boards.greenhouse.io/anthropic/jobs/5205704008\\n[E46](https://job-boards.greenhouse.io/anthropic/jobs/5247640008) Product Manager, Claude Code Model Performance job · https://job-boards.greenhouse.io/anthropic/jobs/5247640008\\n[E47](https://job-boards.greenhouse.io/anthropic/jobs/5247156008) Engineering Manager, Safeguards Data Infrastructure job · https://job-boards.greenhouse.io/anthropic/jobs/5247156008\\n[E48](https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.25.0) anthropics/anthropic-sdk-csharp Anthropic-v12.25.0 · https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.25.0\\n[E49](https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.36.0) anthropics/anthropic-sdk-java v2.36.0 · https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.36.0\\n[E50](https://www.anthropic.com/research/measuring-agent-autonomy) Measuring Agent Autonomy post · https://www.anthropic.com/research/measuring-agent-autonomy\\n[E51](https://www.anthropic.com/research/values-wild) Values Wild post · https://www.anthropic.com/research/values-wild\\n[E52](https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic) How AI Is Transforming Work At Anthropic post · https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic\\n[E53](https://www.anthropic.com/research/assistant-axis) Assistant Axis post · https://www.anthropic.com/research/assistant-axis\\n[E54](https://www.anthropic.com/research/introspection) Introspection post · https://www.anthropic.com/research/introspection\\n[E55](https://www.anthropic.com/research/estimating-productivity-gains) Estimating Productivity Gains post · https://www.anthropic.com/research/estimating-productivity-gains\\n[E56](https://www.anthropic.com/research/claude-personal-guidance) Claude Personal Guidance post · https://www.anthropic.com/research/claude-personal-guidance\\n[E57](https://job-boards.greenhouse.io/anthropic/jobs/5201715008) Applied AI Architect Lead, EMEA Commercial job · https://job-boards.greenhouse.io/anthropic/jobs/5201715008\\n[E58](https://www.anthropic.com/research/emergent-misalignment-reward-hacking) Emergent Misalignment Reward Hacking post · https://www.anthropic.com/research/emergent-misalignment-reward-hacking\\n[E59](https://www.anthropic.com/research/emotion-concepts-function) Emotion Concepts Function post · https://www.anthropic.com/research/emotion-concepts-function\\n[E60](https://www.anthropic.com/research/next-generation-constitutional-classifiers) Next Generation Constitutional Classifiers post · https://www.anthropic.com/research/next-generation-constitutional-classifiers\",\n  \"cites\": [\n    \"P1\", \"P2\", \"P3\", \"P4\", \"P5\", \"P6\", \"P7\", \"P8\", \"P9\", \"P10\",\n    \"P11\", \"P12\", \"P13\", \"P14\", \"P15\", \"P17\", \"P18\", \"P19\", \"P20\", \"P21\",\n    \"P22\", \"P24\", \"P25\", \"P26\", \"P27\", \"P28\",\n    \"E1\", \"E3\", \"E4\", \"E5\", \"E6\", \"E7\", \"E8\", \"E9\", \"E10\", \"E11\",\n    \"E12\", \"E13\", \"E14\", \"E15\", \"E16\", \"E17\", \"E18\", \"E19\", \"E20\", \"E21\",\n    \"E22\", \"E23\", \"E24\", \"E25\", \"E26\", \"E27\", \"E28\", \"E29\", \"E30\", \"E31\",\n    \"E32\", \"E33\", \"E34\", \"E35\", \"E36\", \"E37\", \"E38\", \"E39\", \"E40\", \"E41\",\n    \"E42\", \"E43\", \"E44\", \"E45\", \"E46\", \"E47\", \"E48\", \"E49\", \"E50\", \"E51\",\n    \"E52\", \"E53\", \"E54\", \"E55\", \"E56\", \"E57\", \"E58\", \"E59\", \"E60\"\n  ]\n}\n```","generated_at":"2026-06-09T03:07:05.175+00:00","citations":[{"url":"https://www.anthropic.com/news/anthropic-amazon-compute","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/higher-limits-spacex","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/anthropic-nec","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/theo-hourmouzis-general-manager-australia-new-zealand","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/enterprise-ai-services-company","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5237973008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5222289008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://www.anthropic.com/engineering/managed-agents","path":null,"label":"anthropic.com/engineering","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.168","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.168","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.93","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://www.anthropic.com/news/claude-opus-4-7","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","path":null,"label":"www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","type":"external"},{"url":"https://www.anthropic.com/research/glasswing-initial-update","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/donating-open-source-petri","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/teaching-claude-why","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/natural-language-autoencoders","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/next-generation-constitutional-classifiers","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5247156008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5201715008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5235692008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5205738008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5208289008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5124623008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/4980335008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5205704008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5205505008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5247640008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://www.anthropic.com/engineering/april-23-postmortem","path":null,"label":"anthropic.com/engineering","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5248946008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5248494008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.167","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.166","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.92","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.167","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.166","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.140","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.139","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.138","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.1","path":null,"label":"anthropics/anthropic-sdk-python","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.107.0","path":null,"label":"anthropics/anthropic-sdk-python","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.106.0","path":null,"label":"anthropics/anthropic-sdk-python","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.1","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.1","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.102.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/aws-sdk-v0.4.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/foundry-sdk-v0.3.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/bedrock-sdk-v0.30.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/vertex-sdk-v0.17.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-typescript/releases/tag/sdk-v0.101.0","path":null,"label":"anthropics/anthropic-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.38.0","path":null,"label":"anthropics/anthropic-sdk-java","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.37.0","path":null,"label":"anthropics/anthropic-sdk-java","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-java/releases/tag/v2.36.0","path":null,"label":"anthropics/anthropic-sdk-java","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.48.0","path":null,"label":"anthropics/anthropic-sdk-go","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-go/releases/tag/v1.47.0","path":null,"label":"anthropics/anthropic-sdk-go","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.46.0","path":null,"label":"anthropics/anthropic-sdk-ruby","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-ruby/releases/tag/v1.45.0","path":null,"label":"anthropics/anthropic-sdk-ruby","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.27.0","path":null,"label":"anthropics/anthropic-sdk-csharp","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.26.0","path":null,"label":"anthropics/anthropic-sdk-csharp","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-csharp/releases/tag/Anthropic-v12.25.0","path":null,"label":"anthropics/anthropic-sdk-csharp","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.27.0","path":null,"label":"anthropics/anthropic-sdk-php","type":"external"},{"url":"https://github.com/anthropics/anthropic-sdk-php/releases/tag/v0.26.0","path":null,"label":"anthropics/anthropic-sdk-php","type":"external"},{"url":"https://github.com/anthropics/buffa/releases/tag/v0.5.0","path":null,"label":"anthropics/buffa","type":"external"},{"url":"https://assets.anthropic.com/m/ec212e6566a0d47/original/Disrupting-the-first-reported-AI-orchestrated-cyber-espionage-campaign.pdf","path":null,"label":"assets.anthropic.com/m","type":"external"},{"url":"https://www.anthropic.com/research/2028-ai-leadership","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/news/election-safeguards-update","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/research/persona-selection-model","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/news/claude-for-creative-work","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/research/anthropic-institute-agenda","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/coding-agents-social-sciences","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/measuring-agent-autonomy","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/values-wild","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/emergent-misalignment-reward-hacking","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/assistant-axis","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/introspection","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/estimating-productivity-gains","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/emotion-concepts-function","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/claude-personal-guidance","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/engineering/how-we-contain-claude","path":null,"label":"anthropic.com/engineering","type":"external"},{"url":"https://www.anthropic.com/research/agents-in-biology","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/making-claude-a-chemist","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/engineering/infrastructure-noise","path":null,"label":"anthropic.com/engineering","type":"external"},{"url":"https://www.anthropic.com/news/the-long-term-benefit-trust","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/research/next-generation-constitutional-classifiers","path":null,"label":"anthropic.com/research","type":"external"}],"provenance":{"provider":"deepseek","model":"deepseek-v4-pro","workflow":"onlylabs-deepagents-analysis-v3","agent":"deepagents"},"evidence":{"total":88,"pages":28,"events":140,"web":0,"signal_desks":{"forks":0,"repos":0,"hiring":13,"talking":13,"releases":34},"data_radar_lanes":{"data":1,"evals":0,"safety":1,"product":5,"infrastructure":4},"data_radar_matches":9}},{"org_slug":"zhipu","url":"https://onlylabs.fyi/analysis/zhipu","json_url":"https://onlylabs.fyi/analysis/zhipu/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/zhipu/evidence.json","dossier_url":"https://onlylabs.fyi/labs/zhipu","org":{"slug":"zhipu","name":"Zhipu AI (GLM)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://z.ai"},"title":"Zhipu AI (GLM) analysis","summary":"Zhipu AI (GLM) is shipping a broad, fast-moving family of open-weight GLM models across text, vision, OCR, speech, and image generation, releasing point versions at high cadence (GLM-4.5 through GLM-5/5.1 plus specialized variants) and backing them with first-party SDKs. The standout signal is reach: its OCR and Flash models are pulling millions of monthly Hugging Face downloads, and its older ChatGLM line remains…","markdown":"## Thesis\n[Zhipu AI (GLM)](https://z.ai) is shipping a broad, fast-moving family of open-weight GLM models across text, vision, OCR, speech, and image generation, releasing point versions at high cadence (GLM-4.5 through GLM-5/5.1 plus specialized variants) and backing them with first-party SDKs. The standout signal is reach: its OCR and Flash models are pulling millions of monthly Hugging Face downloads, and its older ChatGLM line remains its most-starred work on GitHub.\n\n## Shipping\nZhipu's most-downloaded model by a wide margin is [`zai-org/GLM-OCR`](https://huggingface.co/zai-org/GLM-OCR) at **4.25M** 30-day downloads (1,815 likes), a compact ~1.3B-param model — and it is on a tight release tempo, with five tagged GitHub releases from [v0.1.1](https://github.com/zai-org/GLM-OCR/releases/tag/v0.1.1) through [v0.1.5](https://github.com/zai-org/GLM-OCR/releases/tag/v0.1.5). Next is [`zai-org/GLM-4.7-Flash`](https://huggingface.co/zai-org/GLM-4.7-Flash) at **1.15M** downloads (~31B params), a smaller/faster tier sitting alongside the full [`zai-org/GLM-4.7`](https://huggingface.co/zai-org/GLM-4.7) (66,802 downloads, ~358B params).\n\nThe flagship frontier line is visibly active: [`zai-org/GLM-5`](https://huggingface.co/zai-org/GLM-5) (97,587 downloads, 2,093 likes — the most-liked model in the set) and [`zai-org/GLM-5.1`](https://huggingface.co/zai-org/GLM-5.1) (134,239 downloads, 1,740 likes), both ~754B params. The GLM-4.5 generation remains heavily pulled too — [`zai-org/GLM-4.5-Air`](https://huggingface.co/zai-org/GLM-4.5-Air) (353,329 downloads, ~110B) and [`zai-org/GLM-4.5`](https://huggingface.co/zai-org/GLM-4.5) (185,074 downloads, ~358B).\n\nModality coverage is wide:\n- **Vision/multimodal:** [`zai-org/GLM-4.1V-9B-Thinking`](https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking) (503,155 downloads), [`zai-org/GLM-4.5V`](https://huggingface.co/zai-org/GLM-4.5V) (177,149), [`zai-org/GLM-4.6V-Flash`](https://huggingface.co/zai-org/GLM-4.6V-Flash) (56,428), [`zai-org/GLM-4.6V`](https://huggingface.co/zai-org/GLM-4.6V) (3,489).\n- **Speech:** [`zai-org/GLM-ASR-Nano-2512`](https://huggingface.co/zai-org/GLM-ASR-Nano-2512) (129,077 downloads, ~2.3B params).\n- **Image generation:** [`zai-org/GLM-Image`](https://huggingface.co/zai-org/GLM-Image) (7,563 downloads, 1,073 likes).\n- **Other:** [`zai-org/GLM-4.6`](https://huggingface.co/zai-org/GLM-4.6) (14,606) and the experimental [`zai-org/Glyph`](https://huggingface.co/zai-org/Glyph) (1,581).\n\nOn tooling, Zhipu maintains first-party SDKs with recent releases: [z-ai-sdk-python v0.2.2](https://github.com/zai-org/z-ai-sdk-python/releases/tag/v0.2.2) and [z-ai-sdk-java 0.3.0](https://github.com/zai-org/z-ai-sdk-java/releases/tag/0.3.0).\n\n## Research themes\nNo first-party writing captured yet.\n\n## Hiring & scaling\nNo careers data captured yet.\n\n## Traction highlights\n- **Hacker News:** [`zai-org/GLM-OCR`](https://github.com/zai-org/GLM-OCR) drew **302 points / 75 comments** — by far the lab's strongest HN thread. The agent-automation repo [`zai-org/Open-AutoGLM`](https://github.com/zai-org/Open-AutoGLM) registered only 2 points / 0 comments.\n- **Most-starred repos:** the legacy chat line dominates — [`ChatGLM-6B`](https://github.com/zai-org/ChatGLM-6B) (41,051 stars), [`Open-AutoGLM`](https://github.com/zai-org/Open-AutoGLM) (25,456), [`ChatGLM2-6B`](https://github.com/zai-org/ChatGLM2-6B) (15,571), [`ChatGLM3`](https://github.com/zai-org/ChatGLM3) (13,680), and [`CogVideo`](https://github.com/zai-org/CogVideo) (12,767). The current [`GLM-OCR`](https://github.com/zai-org/GLM-OCR) repo already sits at 6,918 stars and [`GLM-4.5`](https://github.com/zai-org/GLM-4.5) at 4,351.\n- **Most-downloaded models:** [`GLM-OCR`](https://huggingface.co/zai-org/GLM-OCR) (4.25M), [`GLM-4.7-Flash`](https://huggingface.co/zai-org/GLM-4.7-Flash) (1.15M), and [`GLM-4.1V-9B-Thinking`](https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking) (503K) lead 30-day pulls.\n\n## Sources\n- Homepage: https://z.ai\n- https://huggingface.co/zai-org/GLM-OCR\n- https://huggingface.co/zai-org/GLM-4.7-Flash\n- https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking\n- https://huggingface.co/zai-org/GLM-4.5-Air\n- https://huggingface.co/zai-org/GLM-5\n- https://huggingface.co/zai-org/GLM-5.1\n- https://huggingface.co/zai-org/GLM-ASR-Nano-2512\n- https://huggingface.co/zai-org/GLM-Image\n- https://github.com/zai-org/GLM-OCR (and releases v0.1.1–v0.1.5)\n- https://github.com/zai-org/ChatGLM-6B\n- https://github.com/zai-org/Open-AutoGLM\n- https://github.com/zai-org/z-ai-sdk-python/releases/tag/v0.2.2\n- https://github.com/zai-org/z-ai-sdk-java/releases/tag/0.3.0","generated_at":"2026-06-08T15:59:10.224+00:00","citations":[{"url":"https://z.ai","path":null,"label":"z.ai","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-OCR","path":null,"label":"zai-org/GLM-OCR","type":"external"},{"url":"https://github.com/zai-org/GLM-OCR/releases/tag/v0.1.1","path":null,"label":"zai-org/GLM-OCR","type":"external"},{"url":"https://github.com/zai-org/GLM-OCR/releases/tag/v0.1.5","path":null,"label":"zai-org/GLM-OCR","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.7-Flash","path":null,"label":"zai-org/GLM-4.7-Flash","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.7","path":null,"label":"zai-org/GLM-4.7","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-5","path":null,"label":"zai-org/GLM-5","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-5.1","path":null,"label":"zai-org/GLM-5.1","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.5-Air","path":null,"label":"zai-org/GLM-4.5-Air","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.5","path":null,"label":"zai-org/GLM-4.5","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.1V-9B-Thinking","path":null,"label":"zai-org/GLM-4.1V-9B-Thinking","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.5V","path":null,"label":"zai-org/GLM-4.5V","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.6V-Flash","path":null,"label":"zai-org/GLM-4.6V-Flash","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.6V","path":null,"label":"zai-org/GLM-4.6V","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-ASR-Nano-2512","path":null,"label":"zai-org/GLM-ASR-Nano-2512","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-Image","path":null,"label":"zai-org/GLM-Image","type":"external"},{"url":"https://huggingface.co/zai-org/GLM-4.6","path":null,"label":"zai-org/GLM-4.6","type":"external"},{"url":"https://huggingface.co/zai-org/Glyph","path":null,"label":"zai-org/Glyph","type":"external"},{"url":"https://github.com/zai-org/z-ai-sdk-python/releases/tag/v0.2.2","path":null,"label":"zai-org/z-ai-sdk-python","type":"external"},{"url":"https://github.com/zai-org/z-ai-sdk-java/releases/tag/0.3.0","path":null,"label":"zai-org/z-ai-sdk-java","type":"external"},{"url":"https://github.com/zai-org/GLM-OCR","path":null,"label":"zai-org/GLM-OCR","type":"external"},{"url":"https://github.com/zai-org/Open-AutoGLM","path":null,"label":"zai-org/Open-AutoGLM","type":"external"},{"url":"https://github.com/zai-org/ChatGLM-6B","path":null,"label":"zai-org/ChatGLM-6B","type":"external"},{"url":"https://github.com/zai-org/ChatGLM2-6B","path":null,"label":"zai-org/ChatGLM2-6B","type":"external"},{"url":"https://github.com/zai-org/ChatGLM3","path":null,"label":"zai-org/ChatGLM3","type":"external"},{"url":"https://github.com/zai-org/CogVideo","path":null,"label":"zai-org/CogVideo","type":"external"},{"url":"https://github.com/zai-org/GLM-4.5","path":null,"label":"zai-org/GLM-4.5","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"xai","url":"https://onlylabs.fyi/analysis/xai","json_url":"https://onlylabs.fyi/analysis/xai/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/xai/evidence.json","dossier_url":"https://onlylabs.fyi/labs/xai","org":{"slug":"xai","name":"xAI","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://x.ai/"},"title":"xAI analysis","summary":"xAI is in a productize-and-distribute phase: its frontier work (Grok) lives behind the API while the public footprint is dominated by developer tooling and open-weight artifacts of prior generations. The xai-sdk-python is shipping rapidly (five releases tracked, through v1.15.0), and the company has open-sourced both the Grok-1/Grok-2 weights and, notably, the X recommendation algorithm — signaling tight integration…","markdown":"## Thesis\n\nxAI is in a productize-and-distribute phase: its frontier work (Grok) lives behind the API while the public footprint is dominated by developer tooling and open-weight artifacts of prior generations. The [`xai-sdk-python`](https://github.com/xai-org/xai-sdk-python) is shipping rapidly (five releases tracked, through [v1.15.0](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.15.0)), and the company has open-sourced both the Grok-1/Grok-2 weights and, notably, the [X recommendation algorithm](https://github.com/xai-org/x-algorithm) — signaling tight integration with the broader X/Musk ecosystem.\n\n## Shipping\n\n- **Open-weight models.** [`xai-org/grok-2`](https://huggingface.co/xai-org/grok-2) leads on usage with **48,070** 30-day downloads and **1,095** likes, while the older [`xai-org/grok-1`](https://huggingface.co/xai-org/grok-1) draws only **287** downloads but a larger **2,414** likes — the like/download split suggests grok-1 is a landmark release people bookmark, whereas grok-2 is the one actually being pulled.\n- **SDK cadence.** The Python SDK is the most actively released artifact: [v1.15.0](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.15.0), [v1.14.0](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.14.0), [v1.13.0](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.13.0), [v1.12.2](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.12.2), and [v1.12.1](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.12.1). The underlying [`xai-proto`](https://github.com/xai-org/xai-proto) hit [v1.0.0](https://github.com/xai-org/xai-proto/releases/tag/v1.0.0).\n- **Repos.** [`grok-1`](https://github.com/xai-org/grok-1) is the flagship at **51,681** stars; [`x-algorithm`](https://github.com/xai-org/x-algorithm) at **26,087**; the published [`grok-prompts`](https://github.com/xai-org/grok-prompts) at **4,147**. Developer-facing repos are smaller: [`xai-sdk-python`](https://github.com/xai-org/xai-sdk-python) **471**, [`xai-cookbook`](https://github.com/xai-org/xai-cookbook) **443**, and [`xai-proto`](https://github.com/xai-org/xai-proto) **122**.\n\n## Research themes\n\nNo first-party writing captured yet.\n\n## Hiring & scaling\n\nThe open roles point at three converging investments. **Infrastructure / compute** is explicit in the Memphis, TN footprint — Sr. Software Engineer (Data Center Automation) and a Memphis Member of Technical Staff — consistent with operating large training clusters. **Frontier model work** shows up in London via Member of Technical Staff – Reasoning and Member of Technical Staff – Model Training. **Product surface and distribution** are heavily staffed in London: Mobile Android Engineer, Mobile iOS Engineer, Backend Engineer – API, Backend Engineer, Member of Technical Staff – Sandbox Service, and Member of Technical Staff – X Money (tying Grok to X's payments push). The presence of Legal Operations Analyst (Singapore), Legal Counsel (London), and tutor roles (Tutor – Competition Math; AI Healthcare and Administration Tutor) signals both geographic/regulatory expansion and continued investment in domain-specific RLHF/eval data.\n\n## Traction highlights\n\n- **Hacker News:** the open-sourcing of [`x-algorithm`](https://github.com/xai-org/x-algorithm) drew a thread at **125 points / 65 comments** — the only HN item captured.\n- **Most-starred repo:** [`grok-1`](https://github.com/xai-org/grok-1) at **51,681** stars, with [`x-algorithm`](https://github.com/xai-org/x-algorithm) (**26,087**) a strong second.\n- **Most-downloaded model:** [`xai-org/grok-2`](https://huggingface.co/xai-org/grok-2) at **48,070** 30-day downloads; [`grok-1`](https://huggingface.co/xai-org/grok-1) holds the most likes (**2,414**).\n\n## Sources\n\n- [xAI homepage](https://x.ai/)\n- [xai-org/grok-2 (Hugging Face)](https://huggingface.co/xai-org/grok-2)\n- [xai-org/grok-1 (Hugging Face)](https://huggingface.co/xai-org/grok-1)\n- [xai-org/grok-1 (GitHub)](https://github.com/xai-org/grok-1)\n- [xai-org/x-algorithm (GitHub)](https://github.com/xai-org/x-algorithm)\n- [xai-org/grok-prompts (GitHub)](https://github.com/xai-org/grok-prompts)\n- [xai-org/xai-sdk-python (GitHub)](https://github.com/xai-org/xai-sdk-python)\n- [xai-org/xai-cookbook (GitHub)](https://github.com/xai-org/xai-cookbook)\n- [xai-org/xai-proto (GitHub)](https://github.com/xai-org/xai-proto)\n- [xai-sdk-python v1.15.0 release](https://github.com/xai-org/xai-sdk-python/releases/tag/v1.15.0)\n- [xai-proto v1.0.0 release](https://github.com/xai-org/xai-proto/releases/tag/v1.0.0)","generated_at":"2026-06-08T15:59:10.113+00:00","citations":[{"url":"https://github.com/xai-org/xai-sdk-python","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/xai-sdk-python/releases/tag/v1.15.0","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/x-algorithm","path":null,"label":"xai-org/x-algorithm","type":"external"},{"url":"https://huggingface.co/xai-org/grok-2","path":null,"label":"xai-org/grok-2","type":"external"},{"url":"https://huggingface.co/xai-org/grok-1","path":null,"label":"xai-org/grok-1","type":"external"},{"url":"https://github.com/xai-org/xai-sdk-python/releases/tag/v1.14.0","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/xai-sdk-python/releases/tag/v1.13.0","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/xai-sdk-python/releases/tag/v1.12.2","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/xai-sdk-python/releases/tag/v1.12.1","path":null,"label":"xai-org/xai-sdk-python","type":"external"},{"url":"https://github.com/xai-org/xai-proto","path":null,"label":"xai-org/xai-proto","type":"external"},{"url":"https://github.com/xai-org/xai-proto/releases/tag/v1.0.0","path":null,"label":"xai-org/xai-proto","type":"external"},{"url":"https://github.com/xai-org/grok-1","path":null,"label":"xai-org/grok-1","type":"external"},{"url":"https://github.com/xai-org/grok-prompts","path":null,"label":"xai-org/grok-prompts","type":"external"},{"url":"https://github.com/xai-org/xai-cookbook","path":null,"label":"xai-org/xai-cookbook","type":"external"},{"url":"https://x.ai/","path":null,"label":"x.ai","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"tencent-hunyuan","url":"https://onlylabs.fyi/analysis/tencent-hunyuan","json_url":"https://onlylabs.fyi/analysis/tencent-hunyuan/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/tencent-hunyuan/evidence.json","dossier_url":"https://onlylabs.fyi/labs/tencent-hunyuan","org":{"slug":"tencent-hunyuan","name":"Tencent Hunyuan","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://hunyuan.tencent.com/"},"title":"Tencent Hunyuan analysis","summary":"Tencent Hunyuan is running broad on open-weight generative media — its public footprint skews heavily toward 3D, video, image, and world-model generation rather than chat LLMs. The most-downloaded asset is a frontier-scale ~298B-param model (tencent/Hy3-preview, 90k downloads/30d), but the highest-starred surface area on GitHub is its visual-generation stack (Hunyuan3D, HunyuanVideo). It is also pushing into newer…","markdown":"## Thesis\n[Tencent Hunyuan](https://hunyuan.tencent.com/) is running broad on open-weight generative media — its public footprint skews heavily toward 3D, video, image, and world-model generation rather than chat LLMs. The most-downloaded asset is a frontier-scale ~298B-param model ([tencent/Hy3-preview](https://huggingface.co/tencent/Hy3-preview), 90k downloads/30d), but the highest-starred surface area on GitHub is its visual-generation stack (Hunyuan3D, HunyuanVideo). It is also pushing into newer modalities — machine translation, audio chat, embodied/robotics, GUI agents, and persistent world models.\n\n## Shipping\nOn Hugging Face, the headline model is [tencent/Hy3-preview](https://huggingface.co/tencent/Hy3-preview) (~298.8B params, 90,043 downloads/30d, 271 likes), accompanied by its base variant [tencent/Hy3-preview-Base](https://huggingface.co/tencent/Hy3-preview-Base) (305 downloads, 24 likes). A full machine-translation line is live — [tencent/Hy-MT2-1.8B](https://huggingface.co/tencent/Hy-MT2-1.8B) (22,884 downloads, 1,101 likes — the most-liked model in the set), [tencent/Hy-MT2-7B](https://huggingface.co/tencent/Hy-MT2-7B) (12,990 downloads), the MoE [tencent/Hy-MT2-30B-A3B](https://huggingface.co/tencent/Hy-MT2-30B-A3B) (6,249 downloads, 453 likes), plus aggressively quantized MT1.5 builds at 2-bit ([tencent/Hy-MT1.5-1.8B-2bit](https://huggingface.co/tencent/Hy-MT1.5-1.8B-2bit), 564 downloads) and 1.25-bit ([tencent/Hy-MT1.5-1.8B-1.25bit](https://huggingface.co/tencent/Hy-MT1.5-1.8B-1.25bit), 332 downloads). Beyond text: [tencent/Covo-Audio-Chat](https://huggingface.co/tencent/Covo-Audio-Chat) (14,760 downloads), the world model [tencent/HY-World-2.0](https://huggingface.co/tencent/HY-World-2.0) (3,716 downloads, 665 likes), image generation via [tencent/HunyuanImage-3.0-Instruct-Distil](https://huggingface.co/tencent/HunyuanImage-3.0-Instruct-Distil) (2,670 downloads), the [tencent/Penguin-VL-2B](https://huggingface.co/tencent/Penguin-VL-2B) / [tencent/Penguin-Encoder](https://huggingface.co/tencent/Penguin-Encoder) vision pair, embodied models [tencent/HY-Embodied-0.5](https://huggingface.co/tencent/HY-Embodied-0.5) (908 likes against just 736 downloads) and [tencent/HY-Embodied-0.5-X](https://huggingface.co/tencent/HY-Embodied-0.5-X), and a GUI agent model [tencent/POINTS-GUI-G](https://huggingface.co/tencent/POINTS-GUI-G).\n\nOn GitHub the visual-generation stack dominates: [Hunyuan3D-2](https://github.com/Tencent-Hunyuan/Hunyuan3D-2) (13,900 stars) leads, followed by [HunyuanVideo](https://github.com/Tencent-Hunyuan/HunyuanVideo) (12,179), [HunyuanVideo-1.5](https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5) (4,467), [HunyuanDiT](https://github.com/Tencent-Hunyuan/HunyuanDiT) (4,293), [Hunyuan3D-2.1](https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1) (3,519), [Hunyuan3D-1](https://github.com/Tencent-Hunyuan/Hunyuan3D-1) (3,475), [HunyuanImage-3.0](https://github.com/Tencent-Hunyuan/HunyuanImage-3.0) (3,117), [HunyuanWorld-1.0](https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0) (2,844), [HY-Motion-1.0](https://github.com/Tencent-Hunyuan/HY-Motion-1.0) (2,384), [HY-World-2.0](https://github.com/Tencent-Hunyuan/HY-World-2.0) (2,191), [HunyuanVideo-Avatar](https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar) (2,111), and [HunyuanVideo-I2V](https://github.com/Tencent-Hunyuan/HunyuanVideo-I2V) (1,824). No formal release entries were captured.\n\n## Research themes\nNo first-party writing captured yet.\n\n## Hiring & scaling\nNo careers data captured yet.\n\n## Traction highlights\n- **Most-starred repos:** [Hunyuan3D-2](https://github.com/Tencent-Hunyuan/Hunyuan3D-2) (13,900 stars) and [HunyuanVideo](https://github.com/Tencent-Hunyuan/HunyuanVideo) (12,179 stars) are the clear anchors.\n- **Most-downloaded models:** [tencent/Hy3-preview](https://huggingface.co/tencent/Hy3-preview) (90,043/30d), [tencent/Hy-MT2-1.8B](https://huggingface.co/tencent/Hy-MT2-1.8B) (22,884), and [tencent/Covo-Audio-Chat](https://huggingface.co/tencent/Covo-Audio-Chat) (14,760).\n- **Most-liked models:** [tencent/Hy-MT2-1.8B](https://huggingface.co/tencent/Hy-MT2-1.8B) (1,101 likes) and [tencent/HY-Embodied-0.5](https://huggingface.co/tencent/HY-Embodied-0.5) (908 likes) — the latter notably high relative to its 736 downloads.\n- **Hacker News:** traction is thin — the top thread is [Tencent-Hunyuan/HY-World-2.0](https://github.com/Tencent-Hunyuan/HY-World-2.0) at just 8 points (0 comments), followed by [HunyuanImage-3.0](https://github.com/Tencent-Hunyuan/HunyuanImage-3.0) and [HY-Motion-1.0](https://github.com/Tencent-Hunyuan/HY-Motion-1.0) at 2 points each, and [HunyuanWorld-Mirror](https://github.com/Tencent-Hunyuan/HunyuanWorld-Mirror) at 1.\n\n## Sources\n- Homepage: https://hunyuan.tencent.com/\n- Models: https://huggingface.co/tencent/Hy3-preview · https://huggingface.co/tencent/Hy-MT2-1.8B · https://huggingface.co/tencent/Hy-MT2-7B · https://huggingface.co/tencent/Hy-MT2-30B-A3B · https://huggingface.co/tencent/Covo-Audio-Chat · https://huggingface.co/tencent/HY-World-2.0 · https://huggingface.co/tencent/HunyuanImage-3.0-Instruct-Distil · https://huggingface.co/tencent/HY-Embodied-0.5 · https://huggingface.co/tencent/POINTS-GUI-G\n- Repos: https://github.com/Tencent-Hunyuan/Hunyuan3D-2 · https://github.com/Tencent-Hunyuan/HunyuanVideo · https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5 · https://github.com/Tencent-Hunyuan/HunyuanImage-3.0 · https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0 · https://github.com/Tencent-Hunyuan/HY-World-2.0\n- Hacker News traction: https://github.com/Tencent-Hunyuan/HY-World-2.0 · https://github.com/Tencent-Hunyuan/HunyuanWorld-Mirror","generated_at":"2026-06-08T15:59:09.94+00:00","citations":[{"url":"https://hunyuan.tencent.com/","path":null,"label":"hunyuan.tencent.com","type":"external"},{"url":"https://huggingface.co/tencent/Hy3-preview","path":null,"label":"tencent/Hy3-preview","type":"external"},{"url":"https://huggingface.co/tencent/Hy3-preview-Base","path":null,"label":"tencent/Hy3-preview-Base","type":"external"},{"url":"https://huggingface.co/tencent/Hy-MT2-1.8B","path":null,"label":"tencent/Hy-MT2-1.8B","type":"external"},{"url":"https://huggingface.co/tencent/Hy-MT2-7B","path":null,"label":"tencent/Hy-MT2-7B","type":"external"},{"url":"https://huggingface.co/tencent/Hy-MT2-30B-A3B","path":null,"label":"tencent/Hy-MT2-30B-A3B","type":"external"},{"url":"https://huggingface.co/tencent/Hy-MT1.5-1.8B-2bit","path":null,"label":"tencent/Hy-MT1.5-1.8B-2bit","type":"external"},{"url":"https://huggingface.co/tencent/Hy-MT1.5-1.8B-1.25bit","path":null,"label":"tencent/Hy-MT1.5-1.8B-1.25bit","type":"external"},{"url":"https://huggingface.co/tencent/Covo-Audio-Chat","path":null,"label":"tencent/Covo-Audio-Chat","type":"external"},{"url":"https://huggingface.co/tencent/HY-World-2.0","path":null,"label":"tencent/HY-World-2.0","type":"external"},{"url":"https://huggingface.co/tencent/HunyuanImage-3.0-Instruct-Distil","path":null,"label":"tencent/HunyuanImage-3.0-Instruct-Distil","type":"external"},{"url":"https://huggingface.co/tencent/Penguin-VL-2B","path":null,"label":"tencent/Penguin-VL-2B","type":"external"},{"url":"https://huggingface.co/tencent/Penguin-Encoder","path":null,"label":"tencent/Penguin-Encoder","type":"external"},{"url":"https://huggingface.co/tencent/HY-Embodied-0.5","path":null,"label":"tencent/HY-Embodied-0.5","type":"external"},{"url":"https://huggingface.co/tencent/HY-Embodied-0.5-X","path":null,"label":"tencent/HY-Embodied-0.5-X","type":"external"},{"url":"https://huggingface.co/tencent/POINTS-GUI-G","path":null,"label":"tencent/POINTS-GUI-G","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/Hunyuan3D-2","path":null,"label":"Tencent-Hunyuan/Hunyuan3D-2","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanVideo","path":null,"label":"Tencent-Hunyuan/HunyuanVideo","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5","path":null,"label":"Tencent-Hunyuan/HunyuanVideo-1.5","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanDiT","path":null,"label":"Tencent-Hunyuan/HunyuanDiT","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1","path":null,"label":"Tencent-Hunyuan/Hunyuan3D-2.1","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/Hunyuan3D-1","path":null,"label":"Tencent-Hunyuan/Hunyuan3D-1","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanImage-3.0","path":null,"label":"Tencent-Hunyuan/HunyuanImage-3.0","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanWorld-1.0","path":null,"label":"Tencent-Hunyuan/HunyuanWorld-1.0","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HY-Motion-1.0","path":null,"label":"Tencent-Hunyuan/HY-Motion-1.0","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HY-World-2.0","path":null,"label":"Tencent-Hunyuan/HY-World-2.0","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar","path":null,"label":"Tencent-Hunyuan/HunyuanVideo-Avatar","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanVideo-I2V","path":null,"label":"Tencent-Hunyuan/HunyuanVideo-I2V","type":"external"},{"url":"https://github.com/Tencent-Hunyuan/HunyuanWorld-Mirror","path":null,"label":"Tencent-Hunyuan/HunyuanWorld-Mirror","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"qwen","url":"https://onlylabs.fyi/analysis/qwen","json_url":"https://onlylabs.fyi/analysis/qwen/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/qwen/evidence.json","dossier_url":"https://onlylabs.fyi/labs/qwen","org":{"slug":"qwen","name":"Qwen (Alibaba Cloud)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://qwen.ai/"},"title":"Qwen (Alibaba Cloud) analysis","summary":"Qwen (Alibaba Cloud) is running one of the most prolific open-weight release cadences in the field, shipping a full ladder of dense and Mixture-of-Experts models — currently the Qwen3.5 and Qwen3.6 generations — across every modality and a parallel agentic coding stack (qwen-code, 25k stars). Adoption is enormous: its current flagship-tier checkpoints each pull millions of Hugging Face downloads in a 30-day window.…","markdown":"## Thesis\n\nQwen (Alibaba Cloud) is running one of the most prolific open-weight release cadences in the field, shipping a full ladder of dense and Mixture-of-Experts models — currently the **Qwen3.5** and **Qwen3.6** generations — across every modality and a parallel **agentic coding** stack ([qwen-code](https://github.com/QwenLM/qwen-code), 25k stars). Adoption is enormous: its current flagship-tier checkpoints each pull millions of Hugging Face downloads in a 30-day window. The lab pairs frontier-scale MoE models (up to **Qwen3.5-397B-A17B**) with a dense small-model line tuned for production and mobile, and backs both with a steady stream of first-party research writing.\n\n## Shipping\n\nAcross modalities, the most-downloaded checkpoints in the context are the small dense **Qwen3.5** instruct models: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) at **9,934,423** 30-day downloads (614 likes) and [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) at **9,277,612** (1,536 likes). The new **Qwen3.6** generation is already pulling heavy traffic — [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) (MoE) at **5,852,936** downloads / 2,038 likes and [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) at **5,541,236** / 1,638 likes.\n\nThe MoE strategy spans sizes: the flagship [Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) (403B params, 17B active; 1,077,681 downloads, 1,504 likes), [Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B) (815,955), and [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) (2,754,795). A dense ladder fills out production and edge use: [Qwen3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) (2,857,230), [Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) (1,841,841), and [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) (2,657,382). Matching `-Base` variants ship for most sizes (e.g. [Qwen3.5-4B-Base](https://huggingface.co/Qwen/Qwen3.5-4B-Base), 205,712 downloads), confirming the standard base-plus-instruct release pattern.\n\nOn GitHub, the lab's top repos are [QwenLM/Qwen3](https://github.com/QwenLM/Qwen3) (**27,290** stars), [QwenLM/qwen-code](https://github.com/QwenLM/qwen-code) (**25,009**), [QwenLM/Qwen](https://github.com/QwenLM/Qwen) (**21,255**), and the multimodal/coding lines [QwenLM/Qwen3-VL](https://github.com/QwenLM/Qwen3-VL) (**19,329**), [QwenLM/Qwen3-Coder](https://github.com/QwenLM/Qwen3-Coder) (**16,601**), and [QwenLM/Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) (**16,491**). Speech and image are active too: [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) (11,800), [Qwen-Image](https://github.com/QwenLM/Qwen-Image) (7,977), and [Qwen3-Omni](https://github.com/QwenLM/Qwen3-Omni) (3,819). Release activity is concentrated in the **qwen-code** agentic CLI, which is on a near-daily nightly cadence — the latest tagged builds run from [v0.17.1](https://github.com/QwenLM/qwen-code/releases/tag/v0.17.1) through nightlies dated 20260604–20260608 — alongside supporting repos [qwen-code-examples v0.1](https://github.com/QwenLM/qwen-code-examples/releases/tag/v0.1) and [qwen-code-action v0.1.1](https://github.com/QwenLM/qwen-code-action/releases/tag/v0.1.1).\n\n## Research themes\n\nQwen's first-party writing traces a consistent arc from early unified multimodal pretraining to today's reasoning and agentic systems:\n\n- **Generalist / unified multimodal models** — the lab's roots: [OFA: Towards Building a One-For-All Model](https://qwenlm.github.io/blog/ofa/), [OFASys: Enabling Multitask Learning with One Line of Code](https://qwenlm.github.io/blog/ofasys/), and [Chinese CLIP](https://qwenlm.github.io/blog/chinese-clip/).\n- **Foundation-model generations** — [Introducing Qwen](https://qwenlm.github.io/blog/qwen/), [Introducing Qwen1.5](https://qwenlm.github.io/blog/qwen1.5/), [Hello Qwen2](https://qwenlm.github.io/blog/qwen2/), and [Qwen2.5: A Party of Foundation Models](https://qwenlm.github.io/blog/qwen2.5/).\n- **Mixture-of-Experts efficiency** — [Qwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters](https://qwenlm.github.io/blog/qwen-moe/), the through-line behind today's A3B/A10B/A17B releases.\n- **Long context** — [Generalizing an LLM from 8k to 1M Context using Qwen-Agent](https://qwenlm.github.io/blog/qwen-agent-2405/) and [Extending the Context Length to 1M Tokens!](https://qwenlm.github.io/blog/qwen2.5-turbo/).\n- **Math & reasoning** — [Introducing Qwen2-Math](https://qwenlm.github.io/blog/qwen2-math/), [Qwen2.5-Math](https://qwenlm.github.io/blog/qwen2.5-math/), [Towards Effective Process Supervision in Mathematical Reasoning](https://qwenlm.github.io/blog/qwen2.5-math-prm/), and the reasoning-first [QwQ: Reflect Deeply on the Boundaries of the Unknown](https://qwenlm.github.io/blog/qwq-32b-preview/) plus its visual counterpart [QVQ: To See the World with Wisdom](https://qwenlm.github.io/blog/qvq-72b-preview/).\n- **Code** — [Code with CodeQwen1.5](https://qwenlm.github.io/blog/codeqwen1.5/) and [Qwen2.5-Coder: Code More, Learn More!](https://qwenlm.github.io/blog/qwen2.5-coder/), where the [Coder family post](https://qwenlm.github.io/blog/qwen2.5-coder-family/) positions Qwen2.5-Coder-32B-Instruct as a SOTA open code model \"matching the coding capabilities of GPT-4o.\"\n- **Audio & vision modalities** — [Qwen-VL](https://qwenlm.github.io/blog/qwen-vl/), [Qwen2-VL: To See the World More Clearly](https://qwenlm.github.io/blog/qwen2-vl/), and [Qwen2-Audio: Chat with Your Voice!](https://qwenlm.github.io/blog/qwen2-audio/).\n\n## Hiring & scaling\n\nThe captured roles are all on the **通义大模型事业部** (Tongyi large-model division), based in **杭州 (Hangzhou)**: algorithm engineer (算法工程师), R&D engineer (研发工程师), and their senior counterparts (高级算法工程师 / 高级研发工程师). The split between algorithm and engineering tracks — each at both standard and senior levels — signals continued investment in both core model research and the production/infra stack behind it, concentrated in a single Hangzhou hub rather than distributed teams.\n\n## Traction highlights\n\n- **Hacker News:** the standout thread is [QwenLM/Qwen3-Omni](https://github.com/QwenLM/Qwen3-Omni) at **571 points / 142 comments**, far ahead of the next — [QwenLM/Qwen](https://github.com/QwenLM/Qwen) (36 points, 51 comments) and [QwenLM/Qwen3-VL-Embedding](https://github.com/QwenLM/Qwen3-VL-Embedding) (11 points). Multimodal/omni work draws the strongest external attention.\n- **Most-starred repos:** [QwenLM/Qwen3](https://github.com/QwenLM/Qwen3) (27,290) and [QwenLM/qwen-code](https://github.com/QwenLM/qwen-code) (25,009) lead, with [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL) (19,329) and [Qwen3-Coder](https://github.com/QwenLM/Qwen3-Coder) (16,601) close behind.\n- **Most-downloaded models:** [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) (9.93M) and [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) (9.28M) dominate 30-day downloads — small dense models carry the bulk of real-world usage, while the 397B-A17B flagship still clears 1M.\n\n## Sources\n\n- Homepage: https://qwen.ai/\n- [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) · [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) · [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) · [Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B) · [Qwen/Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) · [Qwen/Qwen3.5-122B-A10B](https://huggingface.co/Qwen/Qwen3.5-122B-A10B)\n- [QwenLM/Qwen3](https://github.com/QwenLM/Qwen3) · [QwenLM/qwen-code](https://github.com/QwenLM/qwen-code) · [QwenLM/Qwen3-VL](https://github.com/QwenLM/Qwen3-VL) · [QwenLM/Qwen3-Coder](https://github.com/QwenLM/Qwen3-Coder) · [QwenLM/Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) · [QwenLM/Qwen3-Omni](https://github.com/QwenLM/Qwen3-Omni)\n- qwen-code releases: [v0.17.1](https://github.com/QwenLM/qwen-code/releases/tag/v0.17.1) · [qwen-code-action v0.1.1](https://github.com/QwenLM/qwen-code-action/releases/tag/v0.1.1)\n- Blog: [Qwen2.5](https://qwenlm.github.io/blog/qwen2.5/) · [Qwen1.5-MoE](https://qwenlm.github.io/blog/qwen-moe/) · [Extending Context to 1M Tokens](https://qwenlm.github.io/blog/qwen2.5-turbo/) · [QwQ](https://qwenlm.github.io/blog/qwq-32b-preview/) · [Qwen2.5-Coder](https://qwenlm.github.io/blog/qwen2.5-coder/) · [Qwen2-VL](https://qwenlm.github.io/blog/qwen2-vl/)","generated_at":"2026-06-08T15:59:09.823+00:00","citations":[{"url":"https://github.com/QwenLM/qwen-code","path":null,"label":"QwenLM/qwen-code","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-4B","path":null,"label":"Qwen/Qwen3.5-4B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-9B","path":null,"label":"Qwen/Qwen3.5-9B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.6-35B-A3B","path":null,"label":"Qwen/Qwen3.6-35B-A3B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.6-27B","path":null,"label":"Qwen/Qwen3.6-27B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-397B-A17B","path":null,"label":"Qwen/Qwen3.5-397B-A17B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-122B-A10B","path":null,"label":"Qwen/Qwen3.5-122B-A10B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-35B-A3B","path":null,"label":"Qwen/Qwen3.5-35B-A3B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-27B","path":null,"label":"Qwen/Qwen3.5-27B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-2B","path":null,"label":"Qwen/Qwen3.5-2B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-0.8B","path":null,"label":"Qwen/Qwen3.5-0.8B","type":"external"},{"url":"https://huggingface.co/Qwen/Qwen3.5-4B-Base","path":null,"label":"Qwen/Qwen3.5-4B-Base","type":"external"},{"url":"https://github.com/QwenLM/Qwen3","path":null,"label":"QwenLM/Qwen3","type":"external"},{"url":"https://github.com/QwenLM/Qwen","path":null,"label":"QwenLM/Qwen","type":"external"},{"url":"https://github.com/QwenLM/Qwen3-VL","path":null,"label":"QwenLM/Qwen3-VL","type":"external"},{"url":"https://github.com/QwenLM/Qwen3-Coder","path":null,"label":"QwenLM/Qwen3-Coder","type":"external"},{"url":"https://github.com/QwenLM/Qwen-Agent","path":null,"label":"QwenLM/Qwen-Agent","type":"external"},{"url":"https://github.com/QwenLM/Qwen3-TTS","path":null,"label":"QwenLM/Qwen3-TTS","type":"external"},{"url":"https://github.com/QwenLM/Qwen-Image","path":null,"label":"QwenLM/Qwen-Image","type":"external"},{"url":"https://github.com/QwenLM/Qwen3-Omni","path":null,"label":"QwenLM/Qwen3-Omni","type":"external"},{"url":"https://github.com/QwenLM/qwen-code/releases/tag/v0.17.1","path":null,"label":"QwenLM/qwen-code","type":"external"},{"url":"https://github.com/QwenLM/qwen-code-examples/releases/tag/v0.1","path":null,"label":"QwenLM/qwen-code-examples","type":"external"},{"url":"https://github.com/QwenLM/qwen-code-action/releases/tag/v0.1.1","path":null,"label":"QwenLM/qwen-code-action","type":"external"},{"url":"https://qwenlm.github.io/blog/ofa/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/ofasys/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/chinese-clip/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen1.5/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen-moe/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen-agent-2405/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5-turbo/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2-math/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5-math/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5-math-prm/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwq-32b-preview/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qvq-72b-preview/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/codeqwen1.5/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5-coder/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2.5-coder-family/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen-vl/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2-vl/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://qwenlm.github.io/blog/qwen2-audio/","path":null,"label":"qwenlm.github.io/blog","type":"external"},{"url":"https://github.com/QwenLM/Qwen3-VL-Embedding","path":null,"label":"QwenLM/Qwen3-VL-Embedding","type":"external"},{"url":"https://qwen.ai/","path":null,"label":"qwen.ai","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"openai","url":"https://onlylabs.fyi/analysis/openai","json_url":"https://onlylabs.fyi/analysis/openai/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/openai/evidence.json","dossier_url":"https://onlylabs.fyi/labs/openai","org":{"slug":"openai","name":"OpenAI","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://openai.com"},"title":"OpenAI analysis","summary":"OpenAI is operating on two fronts at once: a frontier-model release cadence aimed at consumers and developers, and a hard pivot into agentic developer tooling. Its public footprint right now is dominated by Codex, a terminal coding agent shipping near-daily alpha builds, and a wave of GPT-5.x launches (GPT-5.5, GPT-5.4, GPT-5.3-Codex) that top Hacker News. The hiring and infra signals point to scaling compute and…","markdown":"## Thesis\n\nOpenAI is operating on two fronts at once: a frontier-model release cadence aimed at consumers and developers, and a hard pivot into agentic developer tooling. Its public footprint right now is dominated by [Codex](https://github.com/openai/codex), a terminal coding agent shipping near-daily alpha builds, and a wave of GPT-5.x launches ([GPT-5.5](https://openai.com/index/introducing-gpt-5-5), [GPT-5.4](https://openai.com/index/introducing-gpt-5-4), [GPT-5.3-Codex](https://openai.com/index/introducing-gpt-5-3-codex)) that top Hacker News. The hiring and infra signals point to scaling compute and \"long-running agents\" as the next bet.\n\n## Shipping\n\nThe most-shipped artifact in the window is **[openai/codex](https://github.com/openai/codex)** (89,479 stars), with a rapid-fire string of prereleases — `rust-v0.138.0-alpha.1` through [`rust-v0.138.0-alpha.6`](https://github.com/openai/codex/releases/tag/rust-v0.138.0-alpha.6), plus the [`rusty-v8-v149.2.0`](https://github.com/openai/codex/releases/tag/rusty-v8-v149.2.0) toolchain build. SDK releases also moved across languages: [openai-dotnet OpenAI_2.11.0](https://github.com/openai/openai-dotnet/releases/tag/OpenAI_2.11.0), [openai-java v4.39.1](https://github.com/openai/openai-java/releases/tag/v4.39.1), and [openai-ruby v0.66.1](https://github.com/openai/openai-ruby/releases/tag/v0.66.1).\n\nOn Hugging Face, the open-weights footprint is led by the Whisper speech-recognition family: [whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo) (8,558,471 downloads), [whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) (5,389,884), [whisper-base](https://huggingface.co/openai/whisper-base) (4,201,560), [whisper-small](https://huggingface.co/openai/whisper-small) (2,420,545), and [whisper-tiny](https://huggingface.co/openai/whisper-tiny) (1,435,712), alongside [clip-vit-large-patch14-336](https://huggingface.co/openai/clip-vit-large-patch14-336) (3,030,061). Newer to the lineup are the safety-classifier models [gpt-oss-safeguard-20b](https://huggingface.co/openai/gpt-oss-safeguard-20b) (42,123 downloads, 21.5B params) and [gpt-oss-safeguard-120b](https://huggingface.co/openai/gpt-oss-safeguard-120b) (29,589 downloads, 120B params).\n\nTop repos beyond Codex: [openai/whisper](https://github.com/openai/whisper) (102,120 stars), [openai-cookbook](https://github.com/openai/openai-cookbook) (74,067), [gym](https://github.com/openai/gym) (37,213), [CLIP](https://github.com/openai/CLIP) (33,706), [openai-python](https://github.com/openai/openai-python) (30,938), and the agent-stack repos [openai-agents-python](https://github.com/openai/openai-agents-python) (26,982), [swarm](https://github.com/openai/swarm) (21,590), [skills](https://github.com/openai/skills) (21,651), [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) (20,420), and [gpt-oss](https://github.com/openai/gpt-oss) (20,145).\n\n## Research themes\n\nCaptured first-party writing skews toward OpenAI's foundational-era posts plus a recent applied case study. Early themes center on generative models and reinforcement learning — [Generative models](https://openai.com/index/generative-models), [OpenAI Gym Beta](https://openai.com/index/openai-gym-beta), [OpenAI technical goals](https://openai.com/index/openai-technical-goals) (a \"living metric\" for agents across Gym environments), and methods papers like [Weight normalization](https://openai.com/index/weight-normalization) and [Adversarial training methods for semi-supervised text classification](https://openai.com/index/adversarial-training-methods-for-semi-supervised-text-classification). The founding mission and AI-safety framing appear in [Introducing OpenAI](https://openai.com/index/introducing-openai) and [Special projects](https://openai.com/index/special-projects). The lone recent applied piece, [Increasing accuracy of pediatric visit notes](https://openai.com/index/summer-health), shows the deployment angle: using OpenAI models to automate clinical visit notes for Summer Health.\n\n## Hiring & scaling\n\nOpen roles cluster hard around **agents and compute infrastructure**. On the research side, [Researcher: Agent Post-Training, API & Power-Users](https://jobs.ashbyhq.com/openai/a68045e3-1fa1-4dd1-8ef4-e87fac2416eb) describes \"the frontier agents OpenAI ships\" in Codex, ChatGPT, and the API — persistent agents for coding, tool use, and computer use. The product-platform side hires for [Software Engineer, Cloud Agents](https://jobs.ashbyhq.com/openai/f6278b60-dd42-4aa8-a3cd-c105f75ae8ae), building orchestration, sandboxing, and reliability for \"long-running agents in the cloud.\"\n\nCompute scaling shows up as a finance function: [Compute & Infrastructure Accounting Manager](https://jobs.ashbyhq.com/openai/67ef8778-71a1-4a20-ae82-3a3c2ab9c780), covering data centers, rack financing, and colocation — signaling a large, complex infra portfolio. Go-to-market and deployment are expanding internationally with [Technical Deployment Lead - Sydney](https://jobs.ashbyhq.com/openai/26e2f493-6f1c-47e3-a34e-e8d29aa88fbc) (Forward Deployed Engineering) plus AI Deployment Engineering Managers in Singapore, New York, and San Francisco. Other tells: an [Associate General Counsel, Hardware IP](https://jobs.ashbyhq.com/openai/3c944352-a389-47d5-aef7-0b2ba6fd1564) role pointing to a hardware/patent push, safety-ops roles (Analytics Engineer, Safety Systems; Operations Enablement PM, User Safety & Risk Operations), and a [Technical Sourcer, Research](https://jobs.ashbyhq.com/openai/fdf2201f-e2ad-4b4a-9da4-776f64b4fa62) embedded in the research org to recruit top AI researchers.\n\n## Traction highlights\n\nHacker News attention is concentrated on the GPT-5.x line: [Introducing GPT-5.5](https://openai.com/index/introducing-gpt-5-5) (1,580 points, 1,056 comments), [Introducing GPT-5.3-Codex](https://openai.com/index/introducing-gpt-5-3-codex) (1,530 / 605), and [Introducing GPT-5.2](https://openai.com/index/introducing-gpt-5-2) (1,195 / 1,083). A research-flavored thread, [An OpenAI model has disproved a central conjecture in discrete geometry](https://openai.com/index/model-disproves-discrete-geometry-conjecture), also drew heavily (1,429 / 1,055), as did [Introducing ChatGPT Images 2.0](https://openai.com/index/introducing-chatgpt-images-2-0) (1,049 / 975), [Introducing GPT-5.4](https://openai.com/index/introducing-gpt-5-4) (1,019 / 805), and [Codex for (almost) everything](https://openai.com/index/codex-for-almost-everything) (1,001 / 559). On GitHub, the most-starred assets are [openai/whisper](https://github.com/openai/whisper) (102,120) and [openai/codex](https://github.com/openai/codex) (89,479); on Hugging Face, [whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo) leads at 8.56M downloads.\n\n## Sources\n\n- https://github.com/openai/codex/releases/tag/rust-v0.138.0-alpha.6\n- https://github.com/openai/codex/releases/tag/rusty-v8-v149.2.0\n- https://github.com/openai/openai-dotnet/releases/tag/OpenAI_2.11.0\n- https://github.com/openai/whisper\n- https://github.com/openai/codex\n- https://github.com/openai/openai-agents-python\n- https://github.com/openai/gpt-oss\n- https://huggingface.co/openai/whisper-large-v3-turbo\n- https://huggingface.co/openai/whisper-large-v3\n- https://huggingface.co/openai/clip-vit-large-patch14-336\n- https://huggingface.co/openai/gpt-oss-safeguard-120b\n- https://openai.com/index/introducing-gpt-5-5\n- https://openai.com/index/introducing-gpt-5-3-codex\n- https://openai.com/index/model-disproves-discrete-geometry-conjecture\n- https://openai.com/index/codex-for-almost-everything\n- https://openai.com/index/summer-health\n- https://openai.com/index/generative-models\n- https://jobs.ashbyhq.com/openai/a68045e3-1fa1-4dd1-8ef4-e87fac2416eb\n- https://jobs.ashbyhq.com/openai/f6278b60-dd42-4aa8-a3cd-c105f75ae8ae\n- https://jobs.ashbyhq.com/openai/67ef8778-71a1-4a20-ae82-3a3c2ab9c780\n- https://jobs.ashbyhq.com/openai/26e2f493-6f1c-47e3-a34e-e8d29aa88fbc","generated_at":"2026-06-08T15:59:09.71+00:00","citations":[{"url":"https://github.com/openai/codex","path":null,"label":"openai/codex","type":"external"},{"url":"https://openai.com/index/introducing-gpt-5-5","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/introducing-gpt-5-4","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/introducing-gpt-5-3-codex","path":null,"label":"openai.com/index","type":"external"},{"url":"https://github.com/openai/codex/releases/tag/rust-v0.138.0-alpha.6","path":null,"label":"openai/codex","type":"external"},{"url":"https://github.com/openai/codex/releases/tag/rusty-v8-v149.2.0","path":null,"label":"openai/codex","type":"external"},{"url":"https://github.com/openai/openai-dotnet/releases/tag/OpenAI_2.11.0","path":null,"label":"openai/openai-dotnet","type":"external"},{"url":"https://github.com/openai/openai-java/releases/tag/v4.39.1","path":null,"label":"openai/openai-java","type":"external"},{"url":"https://github.com/openai/openai-ruby/releases/tag/v0.66.1","path":null,"label":"openai/openai-ruby","type":"external"},{"url":"https://huggingface.co/openai/whisper-large-v3-turbo","path":null,"label":"openai/whisper-large-v3-turbo","type":"external"},{"url":"https://huggingface.co/openai/whisper-large-v3","path":null,"label":"openai/whisper-large-v3","type":"external"},{"url":"https://huggingface.co/openai/whisper-base","path":null,"label":"openai/whisper-base","type":"external"},{"url":"https://huggingface.co/openai/whisper-small","path":null,"label":"openai/whisper-small","type":"external"},{"url":"https://huggingface.co/openai/whisper-tiny","path":null,"label":"openai/whisper-tiny","type":"external"},{"url":"https://huggingface.co/openai/clip-vit-large-patch14-336","path":null,"label":"openai/clip-vit-large-patch14-336","type":"external"},{"url":"https://huggingface.co/openai/gpt-oss-safeguard-20b","path":null,"label":"openai/gpt-oss-safeguard-20b","type":"external"},{"url":"https://huggingface.co/openai/gpt-oss-safeguard-120b","path":null,"label":"openai/gpt-oss-safeguard-120b","type":"external"},{"url":"https://github.com/openai/whisper","path":null,"label":"openai/whisper","type":"external"},{"url":"https://github.com/openai/openai-cookbook","path":null,"label":"openai/openai-cookbook","type":"external"},{"url":"https://github.com/openai/gym","path":null,"label":"openai/gym","type":"external"},{"url":"https://github.com/openai/CLIP","path":null,"label":"openai/CLIP","type":"external"},{"url":"https://github.com/openai/openai-python","path":null,"label":"openai/openai-python","type":"external"},{"url":"https://github.com/openai/openai-agents-python","path":null,"label":"openai/openai-agents-python","type":"external"},{"url":"https://github.com/openai/swarm","path":null,"label":"openai/swarm","type":"external"},{"url":"https://github.com/openai/skills","path":null,"label":"openai/skills","type":"external"},{"url":"https://github.com/openai/codex-plugin-cc","path":null,"label":"openai/codex-plugin-cc","type":"external"},{"url":"https://github.com/openai/gpt-oss","path":null,"label":"openai/gpt-oss","type":"external"},{"url":"https://openai.com/index/generative-models","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/openai-gym-beta","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/openai-technical-goals","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/weight-normalization","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/adversarial-training-methods-for-semi-supervised-text-classification","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/introducing-openai","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/special-projects","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/summer-health","path":null,"label":"openai.com/index","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/a68045e3-1fa1-4dd1-8ef4-e87fac2416eb","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/f6278b60-dd42-4aa8-a3cd-c105f75ae8ae","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/67ef8778-71a1-4a20-ae82-3a3c2ab9c780","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/26e2f493-6f1c-47e3-a34e-e8d29aa88fbc","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/3c944352-a389-47d5-aef7-0b2ba6fd1564","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://jobs.ashbyhq.com/openai/fdf2201f-e2ad-4b4a-9da4-776f64b4fa62","path":null,"label":"jobs.ashbyhq.com/openai","type":"external"},{"url":"https://openai.com/index/introducing-gpt-5-2","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/model-disproves-discrete-geometry-conjecture","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/introducing-chatgpt-images-2-0","path":null,"label":"openai.com/index","type":"external"},{"url":"https://openai.com/index/codex-for-almost-everything","path":null,"label":"openai.com/index","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"nvidia","url":"https://onlylabs.fyi/analysis/nvidia","json_url":"https://onlylabs.fyi/analysis/nvidia/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/nvidia/evidence.json","dossier_url":"https://onlylabs.fyi/labs/nvidia","org":{"slug":"nvidia","name":"NVIDIA","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.nvidia.com"},"title":"NVIDIA analysis","summary":"NVIDIA is positioning itself as the full-stack supplier of the \"AI factory\" era — selling not just silicon but open models, agent runtimes, and physical-AI foundation models that run on its hardware. The current push centers on three fronts: long-running agents (the Nemotron 3 Ultra family and the NemoClaw agent blueprint), physical/world AI (Cosmos 3 and robotics), and local/personal agents on new hardware (RTX…","markdown":"## Thesis\n\nNVIDIA is positioning itself as the full-stack supplier of the \"AI factory\" era — selling not just silicon but open models, agent runtimes, and physical-AI foundation models that run on its hardware. The current push centers on three fronts: long-running agents (the [Nemotron 3 Ultra](https://blogs.nvidia.com/blog/nvidia-gtc-taipei-computex-2026-news/) family and the [NemoClaw](https://github.com/NVIDIA/NemoClaw) agent blueprint), physical/world AI ([Cosmos 3](https://blogs.nvidia.com/blog/cosmos-3-physical-ai-open-world-foundation-model/) and robotics), and local/personal agents on new hardware (RTX Spark, DGX Spark, Jetson). Nearly all first-party writing in the window is GTC Taipei / COMPUTEX launch and partnership coverage, framing NVIDIA as the infrastructure layer that converts \"energy into tokens.\"\n\n## Shipping\n\nThe flagship open release is **Nemotron 3 Ultra**, an open model built for long-running agents — the [`nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) checkpoint (~560B params) leads the model footprint at **49,784** downloads / 158 likes, with companion Base ([1,059 downloads](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16)) and GenRM reward-model ([413 downloads](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM)) variants. The **Cosmos 3** world-model line ships [`Cosmos3-Super-Text2Image`](https://huggingface.co/nvidia/Cosmos3-Super-Text2Image) (5,075 dl), [`Cosmos3-Super-Image2Video`](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video) (4,515 dl), and the robotics-policy [`Cosmos3-Nano-Policy-DROID`](https://huggingface.co/nvidia/Cosmos3-Nano-Policy-DROID) (4,153 dl). Smaller Nemotron-branded releases cover multimodal and speech — [`Nemotron-Labs-Diffusion-VLM-8B`](https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-VLM-8B) (5,978 dl) and the streaming-ASR [`nemotron-3.5-asr-streaming-0.6b`](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b) (3,439 dl, the most-liked model at 264) — plus a safety classifier, [`Nemotron-3.5-Content-Safety`](https://huggingface.co/nvidia/Nemotron-3.5-Content-Safety) (494 dl).\n\nOn GitHub, the headline repo is [`NVIDIA/NemoClaw`](https://github.com/NVIDIA/NemoClaw) at **21,050 stars** — the open agent blueprint, described in posts as \"an open blueprint for building specialized, long-running agents with a secure runtime and frontier models.\" The training/inference stack remains heavily starred: [`Megatron-LM`](https://github.com/NVIDIA/Megatron-LM) (16,624), [`TensorRT-LLM`](https://github.com/NVIDIA/TensorRT-LLM) (13,825), [`cutlass`](https://github.com/NVIDIA/cutlass) (9,859), and [`nccl`](https://github.com/NVIDIA/nccl) (4,791). Physical-AI and tooling repos round it out: [`cosmos`](https://github.com/NVIDIA/cosmos) (9,677), [`Isaac-GR00T`](https://github.com/NVIDIA/Isaac-GR00T) (7,280), [`warp`](https://github.com/NVIDIA/warp) (6,736), and the LLM red-teaming tool [`garak`](https://github.com/NVIDIA/garak) (8,050). Recent releases are mostly infra/tooling: [`Model-Optimizer 0.45.0rc0`](https://github.com/NVIDIA/Model-Optimizer/releases/tag/0.45.0rc0), [`NeMo-text-processing r1.2.0`](https://github.com/NVIDIA/NeMo-text-processing/releases/tag/r1.2.0), and the front-end component library [`@nvidia-elements/core-v0.2.4`](https://github.com/NVIDIA/elements/releases/tag/%40nvidia-elements/core-v0.2.4).\n\n## Research themes\n\nFirst-party writing clusters into a few clear directions:\n\n- **AI factories as a unit of infrastructure** — the conceptual frame in [\"AI Factories: The New Infrastructure of Intelligence\"](https://blogs.nvidia.com/blog/ai-factories-the-new-infrastructure-of-intelligence/) (converting \"energy into tokens\"; economics defined by tokens/sec, tokens/watt) and the [Vera CPU](https://blogs.nvidia.com/blog/vera-cpu-phoronix/) post on agentic-workload silicon (88 Olympus cores, 1.2TB/s bandwidth).\n- **Long-running and agentic AI** — Nemotron 3 Ultra \"built for long-running agents,\" the [Microsoft unified-stack partnership](https://blogs.nvidia.com/blog/microsoft-build-windows-local-cloud-devices/), and [NemoClaw-based \"autonomous AI engineers\"](https://blogs.nvidia.com/blog/industrial-software-leaders-secure-autonomous-ai-engineers-nemoclaw/) for industrial software.\n- **Physical AI / sim-to-real robotics** — [\"How Cosmos 3 Helps Physical AI Think Before It Acts\"](https://blogs.nvidia.com/blog/cosmos-3-physical-ai-open-world-foundation-model/), the [ICRA sim-to-real paper round-up](https://blogs.nvidia.com/blog/icra-research-robotics-simulation-to-real-world/) (8 of 28 accepted papers), and [CVPR work on grasping, autonomous driving, and agent training at scale](https://blogs.nvidia.com/blog/cvpr-research-grasping-driving-agent-training/).\n- **Local / personal agents** — [RTX Spark and DGX Spark for local agents](https://blogs.nvidia.com/blog/rtx-ai-garage-computex-spark-local-agents/) and [Jetson + NemoClaw at the edge](https://blogs.nvidia.com/blog/jetson-agentic-ai-physical-world/).\n- **Domain foundation models** — the PRAGMA transaction foundation model with Revolut Research, captured both as an [arXiv paper (2604.08649)](https://arxiv.org/pdf/2604.08649) and a [blog explainer](https://blogs.nvidia.com/blog/financial-institutions-transaction-foundation-models/).\n\nA second strand is sovereign-AI / partnership PR — [UK sovereign AI](https://blogs.nvidia.com/blog/uk-sovereign-ai-advancements/), [LG](https://blogs.nvidia.com/blog/nvidia-and-lg-group-ai-factory/) and [Doosan](https://blogs.nvidia.com/blog/nvidia-and-doosan-group-physical-ai/) AI factories, and [Taiwan's Vera Rubin supply chain](https://blogs.nvidia.com/blog/taiwan-ecosystem-ai-infrastructure/) — which reads more as ecosystem/go-to-market than research.\n\n## Hiring & scaling\n\nNo careers data captured yet.\n\n## Traction highlights\n\nOn Hacker News, NVIDIA's open developer tools and agent stack drove the most discussion: [`NVIDIA/warp`](https://github.com/NVIDIA/warp) topped the list at **490 points / 136 comments**, followed by the [`NemoClaw`](https://github.com/NVIDIA/NemoClaw) agent blueprint at **385 points / 261 comments** (the most-commented thread), the [`garak`](https://github.com/NVIDIA/garak) LLM red-teaming tool at **211 points / 62 comments**, and [`NVIDIA/MatX`](https://github.com/NVIDIA/MatX) at **103 points / 79 comments**. The GTC Taipei live-updates post drew only minor HN attention (4 points).\n\nMost-starred repos: [`NemoClaw`](https://github.com/NVIDIA/NemoClaw) (21,050), [`Megatron-LM`](https://github.com/NVIDIA/Megatron-LM) (16,624), and [`TensorRT-LLM`](https://github.com/NVIDIA/TensorRT-LLM) (13,825). Most-downloaded models: [`Nemotron-3-Ultra-550B-A55B-BF16`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) (49,784), [`Nemotron-Labs-Diffusion-VLM-8B`](https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-VLM-8B) (5,978), and [`Cosmos3-Super-Text2Image`](https://huggingface.co/nvidia/Cosmos3-Super-Text2Image) (5,075).\n\n## Sources\n\n- [Nemotron-3-Ultra-550B-A55B-BF16 (Hugging Face)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)\n- [Cosmos3-Super-Text2Image (Hugging Face)](https://huggingface.co/nvidia/Cosmos3-Super-Text2Image)\n- [nemotron-3.5-asr-streaming-0.6b (Hugging Face)](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b)\n- [NVIDIA/NemoClaw (GitHub)](https://github.com/NVIDIA/NemoClaw)\n- [NVIDIA/Megatron-LM (GitHub)](https://github.com/NVIDIA/Megatron-LM)\n- [NVIDIA/TensorRT-LLM (GitHub)](https://github.com/NVIDIA/TensorRT-LLM)\n- [NVIDIA/warp (GitHub)](https://github.com/NVIDIA/warp)\n- [NVIDIA/garak (GitHub)](https://github.com/NVIDIA/garak)\n- [NVIDIA GTC Taipei at COMPUTEX: Live Updates](https://blogs.nvidia.com/blog/nvidia-gtc-taipei-computex-2026-news/)\n- [How Cosmos 3 Helps Physical AI Think Before It Acts](https://blogs.nvidia.com/blog/cosmos-3-physical-ai-open-world-foundation-model/)\n- [AI Factories: The New Infrastructure of Intelligence](https://blogs.nvidia.com/blog/ai-factories-the-new-infrastructure-of-intelligence/)\n- [NVIDIA Vera CPU vs. Competition](https://blogs.nvidia.com/blog/vera-cpu-phoronix/)\n- [NVIDIA Levels Up Local AI Agents Across RTX PCs and DGX Spark](https://blogs.nvidia.com/blog/rtx-ai-garage-computex-spark-local-agents/)\n- [Industrial Software Leaders Build AI Engineers With NemoClaw](https://blogs.nvidia.com/blog/industrial-software-leaders-secure-autonomous-ai-engineers-nemoclaw/)\n- [NVIDIA Research Advances Robotics From Simulation to the Real World (ICRA)](https://blogs.nvidia.com/blog/icra-research-robotics-simulation-to-real-world/)\n- [PRAGMA: Revolut Foundation Model (arXiv 2604.08649)](https://arxiv.org/pdf/2604.08649)\n- [NVIDIA Partners With Microsoft on Unified Stack for Agentic AI](https://blogs.nvidia.com/blog/microsoft-build-windows-local-cloud-devices/)","generated_at":"2026-06-08T15:59:09.594+00:00","citations":[{"url":"https://blogs.nvidia.com/blog/nvidia-gtc-taipei-computex-2026-news/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://github.com/NVIDIA/NemoClaw","path":null,"label":"NVIDIA/NemoClaw","type":"external"},{"url":"https://blogs.nvidia.com/blog/cosmos-3-physical-ai-open-world-foundation-model/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16","path":null,"label":"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16","type":"external"},{"url":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16","path":null,"label":"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16","type":"external"},{"url":"https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM","path":null,"label":"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM","type":"external"},{"url":"https://huggingface.co/nvidia/Cosmos3-Super-Text2Image","path":null,"label":"nvidia/Cosmos3-Super-Text2Image","type":"external"},{"url":"https://huggingface.co/nvidia/Cosmos3-Super-Image2Video","path":null,"label":"nvidia/Cosmos3-Super-Image2Video","type":"external"},{"url":"https://huggingface.co/nvidia/Cosmos3-Nano-Policy-DROID","path":null,"label":"nvidia/Cosmos3-Nano-Policy-DROID","type":"external"},{"url":"https://huggingface.co/nvidia/Nemotron-Labs-Diffusion-VLM-8B","path":null,"label":"nvidia/Nemotron-Labs-Diffusion-VLM-8B","type":"external"},{"url":"https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b","path":null,"label":"nvidia/nemotron-3.5-asr-streaming-0.6b","type":"external"},{"url":"https://huggingface.co/nvidia/Nemotron-3.5-Content-Safety","path":null,"label":"nvidia/Nemotron-3.5-Content-Safety","type":"external"},{"url":"https://github.com/NVIDIA/Megatron-LM","path":null,"label":"NVIDIA/Megatron-LM","type":"external"},{"url":"https://github.com/NVIDIA/TensorRT-LLM","path":null,"label":"NVIDIA/TensorRT-LLM","type":"external"},{"url":"https://github.com/NVIDIA/cutlass","path":null,"label":"NVIDIA/cutlass","type":"external"},{"url":"https://github.com/NVIDIA/nccl","path":null,"label":"NVIDIA/nccl","type":"external"},{"url":"https://github.com/NVIDIA/cosmos","path":null,"label":"NVIDIA/cosmos","type":"external"},{"url":"https://github.com/NVIDIA/Isaac-GR00T","path":null,"label":"NVIDIA/Isaac-GR00T","type":"external"},{"url":"https://github.com/NVIDIA/warp","path":null,"label":"NVIDIA/warp","type":"external"},{"url":"https://github.com/NVIDIA/garak","path":null,"label":"NVIDIA/garak","type":"external"},{"url":"https://github.com/NVIDIA/Model-Optimizer/releases/tag/0.45.0rc0","path":null,"label":"NVIDIA/Model-Optimizer","type":"external"},{"url":"https://github.com/NVIDIA/NeMo-text-processing/releases/tag/r1.2.0","path":null,"label":"NVIDIA/NeMo-text-processing","type":"external"},{"url":"https://github.com/NVIDIA/elements/releases/tag/%40nvidia-elements/core-v0.2.4","path":null,"label":"NVIDIA/elements","type":"external"},{"url":"https://blogs.nvidia.com/blog/ai-factories-the-new-infrastructure-of-intelligence/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/vera-cpu-phoronix/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/microsoft-build-windows-local-cloud-devices/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/industrial-software-leaders-secure-autonomous-ai-engineers-nemoclaw/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/icra-research-robotics-simulation-to-real-world/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/cvpr-research-grasping-driving-agent-training/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/rtx-ai-garage-computex-spark-local-agents/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/jetson-agentic-ai-physical-world/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://arxiv.org/pdf/2604.08649","path":null,"label":"arxiv.org/pdf","type":"external"},{"url":"https://blogs.nvidia.com/blog/financial-institutions-transaction-foundation-models/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/uk-sovereign-ai-advancements/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/nvidia-and-lg-group-ai-factory/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/nvidia-and-doosan-group-physical-ai/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://blogs.nvidia.com/blog/taiwan-ecosystem-ai-infrastructure/","path":null,"label":"blogs.nvidia.com/blog","type":"external"},{"url":"https://github.com/NVIDIA/MatX","path":null,"label":"NVIDIA/MatX","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"moonshot","url":"https://onlylabs.fyi/analysis/moonshot","json_url":"https://onlylabs.fyi/analysis/moonshot/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/moonshot/evidence.json","dossier_url":"https://onlylabs.fyi/labs/moonshot","org":{"slug":"moonshot","name":"Moonshot AI (Kimi)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.moonshot.ai/"},"title":"Moonshot AI (Kimi) analysis","summary":"Moonshot AI (Kimi) is shipping open-weight, trillion-parameter mixture-of-experts frontier models at a fast iteration cadence — the Kimi-K2 line is its flagship, now through K2.5 and K2.6 plus a dedicated K2-Thinking variant. Alongside the weights it is building a full agentic-coding surface (the kimi-cli / kimi-code tools) and publishing efficiency-oriented architecture research (linear attention, attention…","markdown":"## Thesis\n\nMoonshot AI (Kimi) is shipping open-weight, trillion-parameter mixture-of-experts frontier models at a fast iteration cadence — the [Kimi-K2 line](https://github.com/MoonshotAI/Kimi-K2) is its flagship, now through K2.5 and K2.6 plus a dedicated K2-Thinking variant. Alongside the weights it is building a full agentic-coding surface (the `kimi-cli` / `kimi-code` tools) and publishing efficiency-oriented architecture research (linear attention, attention residuals, block attention). It is a Beijing-based lab pairing aggressive open releases with developer tooling.\n\n## Shipping\n\n- **Kimi-K2 family (≈1T-param MoE).** The most-downloaded model is [`moonshotai/Kimi-K2.6`](https://huggingface.co/moonshotai/Kimi-K2.6) at **3,139,192** 30-day downloads (1,417 likes), followed by [`Kimi-K2-Instruct-0905`](https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905) at **2,734,600** and [`Kimi-K2.5`](https://huggingface.co/moonshotai/Kimi-K2.5) at **1,682,758** (the single most-liked card at 2,813 likes). The original [`Kimi-K2-Instruct`](https://huggingface.co/moonshotai/Kimi-K2-Instruct) adds **629,908** downloads, a [`Kimi-K2-Base`](https://huggingface.co/moonshotai/Kimi-K2-Base) (42,722) is published, and a reasoning-focused [`Kimi-K2-Thinking`](https://huggingface.co/moonshotai/Kimi-K2-Thinking) has **163,935** downloads (1,699 likes). The [Kimi-K2 GitHub repo](https://github.com/MoonshotAI/Kimi-K2) is the lab's top repo at **10,839 stars**.\n- **Multimodal (Kimi-VL).** [`Kimi-VL-A3B-Instruct`](https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct) (297,546 downloads) and [`Kimi-VL-A3B-Thinking`](https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking) (134,822), with a [`Kimi-VL-A3B-Thinking-2506`](https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506) refresh and the [`MoonViT-SO-400M`](https://huggingface.co/moonshotai/MoonViT-SO-400M) vision encoder; the [Kimi-VL repo](https://github.com/MoonshotAI/Kimi-VL) has 1,198 stars.\n- **Efficiency / architecture lines.** [`Moonlight-16B-A3B-Instruct`](https://huggingface.co/moonshotai/Moonlight-16B-A3B-Instruct) (101,711 downloads) and [`Kimi-Linear-48B-A3B-Instruct`](https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct) (61,936) — sparse A3B (≈3B active) designs.\n- **Audio & coding.** [`Kimi-Audio-7B-Instruct`](https://huggingface.co/moonshotai/Kimi-Audio-7B-Instruct) (89,403 downloads; [Kimi-Audio repo](https://github.com/MoonshotAI/Kimi-Audio) at 4,645 stars) and the SWE-oriented [`Kimi-Dev-72B`](https://huggingface.co/moonshotai/Kimi-Dev-72B).\n- **Developer tooling, actively versioned.** [`kimi-cli`](https://github.com/MoonshotAI/kimi-cli) (8,916 stars) is shipping rapidly — releases [1.44.0](https://github.com/MoonshotAI/kimi-cli/releases/tag/1.44.0) through [1.47.0](https://github.com/MoonshotAI/kimi-cli/releases/tag/1.47.0) — alongside the [`kimi-code`](https://github.com/MoonshotAI/kimi-code) package (2,041 stars), versioned [0.8.0 → 0.11.0](https://github.com/MoonshotAI/kimi-code/releases/tag/%40moonshot-ai/kimi-code%400.11.0), and a [`walle` v0.1.10](https://github.com/MoonshotAI/walle/releases/tag/v0.1.10) release.\n\n## Research themes\n\nNo first-party writing captured yet. The themes are inferable only from open-source repos: efficiency-oriented attention/architecture research — [Attention-Residuals](https://github.com/MoonshotAI/Attention-Residuals) (3,299 stars), [MoBA](https://github.com/MoonshotAI/MoBA) (block attention, 2,123 stars), [Kimi-Linear](https://github.com/MoonshotAI/Kimi-Linear) (linear attention, 1,399 stars), and the [Kimi-k1.5](https://github.com/MoonshotAI/Kimi-k1.5) reasoning work (3,472 stars).\n\n## Hiring & scaling\n\nAll seven open roles are based in Beijing (北京), signaling a single-hub build-out. The mix is research-and-platform heavy: an algorithm researcher (算法研究员) and algorithm engineer (算法工程师) point to continued model R&D, while backend (后端开发工程师), frontend (前端开发工程师), product manager (产品经理), and designer (设计师) roles indicate investment in a productized, user-facing surface (consistent with the Kimi app and CLI tooling). One operations specialist (运营专员) role rounds out a go-to-market push rather than a pure research lab posture.\n\n## Traction highlights\n\n- **Hacker News:** the lab's architecture research drew the most attention — [Attention-Residuals](https://github.com/MoonshotAI/Attention-Residuals) at **241 points / 34 comments** and [Kimi-Linear](https://github.com/MoonshotAI/Kimi-Linear) at **217 points / 47 comments**; smaller threads appeared for [kimi-cli](https://github.com/MoonshotAI/kimi-cli) (5 pts), [checkpoint-engine](https://github.com/MoonshotAI/checkpoint-engine) (2 pts), and [FlashKDA](https://github.com/MoonshotAI/FlashKDA) (2 pts).\n- **Most-starred repos:** [Kimi-K2](https://github.com/MoonshotAI/Kimi-K2) (10,839), [kimi-cli](https://github.com/MoonshotAI/kimi-cli) (8,916), [Kimi-Audio](https://github.com/MoonshotAI/Kimi-Audio) (4,645).\n- **Most-downloaded models (30-day):** [Kimi-K2.6](https://huggingface.co/moonshotai/Kimi-K2.6) (3.14M), [Kimi-K2-Instruct-0905](https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905) (2.73M), [Kimi-K2.5](https://huggingface.co/moonshotai/Kimi-K2.5) (1.68M).\n\n## Sources\n\n- https://www.moonshot.ai/\n- https://huggingface.co/moonshotai/Kimi-K2.6\n- https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905\n- https://huggingface.co/moonshotai/Kimi-K2.5\n- https://huggingface.co/moonshotai/Kimi-K2-Instruct\n- https://huggingface.co/moonshotai/Kimi-K2-Thinking\n- https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct\n- https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct\n- https://huggingface.co/moonshotai/Kimi-Audio-7B-Instruct\n- https://huggingface.co/moonshotai/Moonlight-16B-A3B-Instruct\n- https://github.com/MoonshotAI/Kimi-K2\n- https://github.com/MoonshotAI/kimi-cli\n- https://github.com/MoonshotAI/kimi-code\n- https://github.com/MoonshotAI/Kimi-Audio\n- https://github.com/MoonshotAI/Attention-Residuals\n- https://github.com/MoonshotAI/Kimi-Linear\n- https://github.com/MoonshotAI/MoBA\n- https://github.com/MoonshotAI/Kimi-k1.5\n- https://github.com/MoonshotAI/Kimi-VL\n- https://github.com/MoonshotAI/walle/releases/tag/v0.1.10","generated_at":"2026-06-08T15:59:09.409+00:00","citations":[{"url":"https://github.com/MoonshotAI/Kimi-K2","path":null,"label":"MoonshotAI/Kimi-K2","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2.6","path":null,"label":"moonshotai/Kimi-K2.6","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2-Instruct-0905","path":null,"label":"moonshotai/Kimi-K2-Instruct-0905","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2.5","path":null,"label":"moonshotai/Kimi-K2.5","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2-Instruct","path":null,"label":"moonshotai/Kimi-K2-Instruct","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2-Base","path":null,"label":"moonshotai/Kimi-K2-Base","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-K2-Thinking","path":null,"label":"moonshotai/Kimi-K2-Thinking","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct","path":null,"label":"moonshotai/Kimi-VL-A3B-Instruct","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking","path":null,"label":"moonshotai/Kimi-VL-A3B-Thinking","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-VL-A3B-Thinking-2506","path":null,"label":"moonshotai/Kimi-VL-A3B-Thinking-2506","type":"external"},{"url":"https://huggingface.co/moonshotai/MoonViT-SO-400M","path":null,"label":"moonshotai/MoonViT-SO-400M","type":"external"},{"url":"https://github.com/MoonshotAI/Kimi-VL","path":null,"label":"MoonshotAI/Kimi-VL","type":"external"},{"url":"https://huggingface.co/moonshotai/Moonlight-16B-A3B-Instruct","path":null,"label":"moonshotai/Moonlight-16B-A3B-Instruct","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct","path":null,"label":"moonshotai/Kimi-Linear-48B-A3B-Instruct","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-Audio-7B-Instruct","path":null,"label":"moonshotai/Kimi-Audio-7B-Instruct","type":"external"},{"url":"https://github.com/MoonshotAI/Kimi-Audio","path":null,"label":"MoonshotAI/Kimi-Audio","type":"external"},{"url":"https://huggingface.co/moonshotai/Kimi-Dev-72B","path":null,"label":"moonshotai/Kimi-Dev-72B","type":"external"},{"url":"https://github.com/MoonshotAI/kimi-cli","path":null,"label":"MoonshotAI/kimi-cli","type":"external"},{"url":"https://github.com/MoonshotAI/kimi-cli/releases/tag/1.44.0","path":null,"label":"MoonshotAI/kimi-cli","type":"external"},{"url":"https://github.com/MoonshotAI/kimi-cli/releases/tag/1.47.0","path":null,"label":"MoonshotAI/kimi-cli","type":"external"},{"url":"https://github.com/MoonshotAI/kimi-code","path":null,"label":"MoonshotAI/kimi-code","type":"external"},{"url":"https://github.com/MoonshotAI/kimi-code/releases/tag/%40moonshot-ai/kimi-code%400.11.0","path":null,"label":"MoonshotAI/kimi-code","type":"external"},{"url":"https://github.com/MoonshotAI/walle/releases/tag/v0.1.10","path":null,"label":"MoonshotAI/walle","type":"external"},{"url":"https://github.com/MoonshotAI/Attention-Residuals","path":null,"label":"MoonshotAI/Attention-Residuals","type":"external"},{"url":"https://github.com/MoonshotAI/MoBA","path":null,"label":"MoonshotAI/MoBA","type":"external"},{"url":"https://github.com/MoonshotAI/Kimi-Linear","path":null,"label":"MoonshotAI/Kimi-Linear","type":"external"},{"url":"https://github.com/MoonshotAI/Kimi-k1.5","path":null,"label":"MoonshotAI/Kimi-k1.5","type":"external"},{"url":"https://github.com/MoonshotAI/checkpoint-engine","path":null,"label":"MoonshotAI/checkpoint-engine","type":"external"},{"url":"https://github.com/MoonshotAI/FlashKDA","path":null,"label":"MoonshotAI/FlashKDA","type":"external"},{"url":"https://www.moonshot.ai/","path":null,"label":"moonshot.ai","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"mistral","url":"https://onlylabs.fyi/analysis/mistral","json_url":"https://onlylabs.fyi/analysis/mistral/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/mistral/evidence.json","dossier_url":"https://onlylabs.fyi/labs/mistral","org":{"slug":"mistral","name":"Mistral AI","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://mistral.ai"},"title":"Mistral AI analysis","summary":"Mistral AI is executing a broad open-weights strategy across every modality and size tier at once: text instruct/reasoning models from 3B up to 128B, a Voxtral audio/speech family (realtime, TTS), and a Devstral coding line. Distribution runs through Hugging Face at serious volume and a full client/tooling stack (mistral-inference, mistral-common, multi-language SDKs). The 2512/2602/2603 release cadence shows rapid,…","markdown":"## Thesis\n\nMistral AI is executing a broad open-weights strategy across every modality and size tier at once: text instruct/reasoning models from 3B up to 128B, a Voxtral audio/speech family (realtime, TTS), and a Devstral coding line. Distribution runs through Hugging Face at serious volume and a full client/tooling stack (`mistral-inference`, `mistral-common`, multi-language SDKs). The 2512/2602/2603 release cadence shows rapid, version-stamped iteration aimed at being the default open frontier alternative.\n\n## Shipping\n\nThe headline asset is speech/audio. [`Voxtral-Mini-4B-Realtime-2602`](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602) leads with **1,187,772** 30-day downloads (871 likes), with siblings [`Voxtral-Mini-3B-2507`](https://huggingface.co/mistralai/Voxtral-Mini-3B-2507) (378,895 downloads), [`Voxtral-Small-24B-2507`](https://huggingface.co/mistralai/Voxtral-Small-24B-2507) (53,674), and a TTS entry [`Voxtral-4B-TTS-2603`](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603) (25,891 downloads, 843 likes).\n\nOn text, the Small tier carries the install base: [`Mistral-Small-3.2-24B-Instruct-2506`](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506) at **585,920** downloads, the newer [`Mistral-Small-4-119B-2603`](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603) (49,665), and the 128B-class [`Mistral-Medium-3.5-128B`](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B) (433,635). The Magistral reasoning line ships [`Magistral-Small-2509`](https://huggingface.co/mistralai/Magistral-Small-2509) (55,566) and [`Magistral-Small-2506`](https://huggingface.co/mistralai/Magistral-Small-2506) (46,228, 608 likes). A new edge-focused Ministral-3 family spans base/instruct/reasoning variants at 3B/8B/14B — e.g. [`Ministral-3-3B-Reasoning-2512`](https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512) (53,201) and [`Ministral-3-8B-Instruct-2512-BF16`](https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16) (51,106). Coding is covered by [`Devstral-Small-2507`](https://huggingface.co/mistralai/Devstral-Small-2507) (33,490 downloads, 365 likes).\n\nThe open-source tooling is meaningful on its own: [`mistral-inference`](https://github.com/mistralai/mistral-inference) (10,814 stars), [`mistral-vibe`](https://github.com/mistralai/mistral-vibe) (4,419 stars, shipping fast at [v2.14.0](https://github.com/mistralai/mistral-vibe/releases/tag/v2.14.0)), [`mistral-finetune`](https://github.com/mistralai/mistral-finetune) (3,093), and [`cookbook`](https://github.com/mistralai/cookbook) (2,259). The plumbing layer — [`mistral-common`](https://github.com/mistralai/mistral-common) ([v1.11.3](https://github.com/mistralai/mistral-common/releases/tag/v1.11.3)) and SDKs [`client-python`](https://github.com/mistralai/client-python) ([v2.4.9](https://github.com/mistralai/client-python/releases/tag/v2.4.9)) and [`client-ts`](https://github.com/mistralai/client-ts) (with dedicated [GCP](https://github.com/mistralai/client-ts/releases/tag/packages/mistralai-gcp/v2.0.0) and [Azure](https://github.com/mistralai/client-ts/releases/tag/packages/mistralai-azure/v2.0.0) v2.0.0 packages) — keeps releasing in lockstep with the models.\n\n## Research themes\n\nNo first-party writing captured yet.\n\n## Hiring & scaling\n\nThe 15 open roles point at commercial deployment and operations rather than core pretraining. The clearest signal is a concentrated \"Applied AI / Forward Deployed\" build-out in Munich — \"Applied AI, Technical Lead, Forward Deployed AI Engineer - Munich\", \"Applied AI, Forward Deployed Machine Learning Engineer - Munich\", and \"Applied AI, Senior/Staff Forward Deployed Machine Learning Engineer - Munich\" — plus an \"AI4Engineering\" vertical in Paris (\"Applied Scientist / Domain Expert, AI4Engineering - EMEA\", \"AI Deployment Strategist, AI4Engineering - EMEA\") and an \"Applied AI Engineer, Site Reliability Engineer - EMEA\". That is a customer-facing, enterprise-deployment motion. The rest is company-scaling overhead: GTM (\"Strategic Account Marketing Manager, APAC\" in Singapore, \"Product Monetisation & Pricing Lead\", \"Solution Operations Manager - Singapore\"), finance/legal (\"Financial Controller\", \"Legal Counsel, Banking / Financing (Project finance)\"), and people/workplace ops across Paris and Palo Alto. Geographically the center of gravity is Paris, with Munich, Singapore (APAC) and Palo Alto (North America) as the expansion edges.\n\n## Traction highlights\n\n- Most-downloaded model: [`Voxtral-Mini-4B-Realtime-2602`](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602) at 1,187,772 30-day downloads, followed by [`Mistral-Small-3.2-24B-Instruct-2506`](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506) (585,920) and [`Mistral-Medium-3.5-128B`](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B) (433,635).\n- Most-starred repo: [`mistral-inference`](https://github.com/mistralai/mistral-inference) (10,814 stars), then [`mistral-vibe`](https://github.com/mistralai/mistral-vibe) (4,419) and [`mistral-finetune`](https://github.com/mistralai/mistral-finetune) (3,093).\n- Hacker News: thin — the only captured thread is [`mistralai/mistral-vibe`](https://github.com/mistralai/mistral-vibe) at just 3 points and 0 comments.\n\n## Sources\n\n- Models: [Voxtral-Mini-4B-Realtime-2602](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602), [Mistral-Small-3.2-24B-Instruct-2506](https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506), [Mistral-Medium-3.5-128B](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B), [Mistral-Small-4-119B-2603](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603), [Magistral-Small-2509](https://huggingface.co/mistralai/Magistral-Small-2509), [Voxtral-4B-TTS-2603](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603), [Devstral-Small-2507](https://huggingface.co/mistralai/Devstral-Small-2507), [Ministral-3-3B-Reasoning-2512](https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512)\n- Repos: [mistral-inference](https://github.com/mistralai/mistral-inference), [mistral-vibe](https://github.com/mistralai/mistral-vibe), [mistral-finetune](https://github.com/mistralai/mistral-finetune), [cookbook](https://github.com/mistralai/cookbook), [mistral-common](https://github.com/mistralai/mistral-common), [client-python](https://github.com/mistralai/client-python), [client-ts](https://github.com/mistralai/client-ts)\n- Releases: [mistral-vibe v2.14.0](https://github.com/mistralai/mistral-vibe/releases/tag/v2.14.0), [mistral-common v1.11.3](https://github.com/mistralai/mistral-common/releases/tag/v1.11.3), [client-python v2.4.9](https://github.com/mistralai/client-python/releases/tag/v2.4.9)\n- Homepage: [mistral.ai](https://mistral.ai)","generated_at":"2026-06-08T15:59:09.295+00:00","citations":[{"url":"https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602","path":null,"label":"mistralai/Voxtral-Mini-4B-Realtime-2602","type":"external"},{"url":"https://huggingface.co/mistralai/Voxtral-Mini-3B-2507","path":null,"label":"mistralai/Voxtral-Mini-3B-2507","type":"external"},{"url":"https://huggingface.co/mistralai/Voxtral-Small-24B-2507","path":null,"label":"mistralai/Voxtral-Small-24B-2507","type":"external"},{"url":"https://huggingface.co/mistralai/Voxtral-4B-TTS-2603","path":null,"label":"mistralai/Voxtral-4B-TTS-2603","type":"external"},{"url":"https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506","path":null,"label":"mistralai/Mistral-Small-3.2-24B-Instruct-2506","type":"external"},{"url":"https://huggingface.co/mistralai/Mistral-Small-4-119B-2603","path":null,"label":"mistralai/Mistral-Small-4-119B-2603","type":"external"},{"url":"https://huggingface.co/mistralai/Mistral-Medium-3.5-128B","path":null,"label":"mistralai/Mistral-Medium-3.5-128B","type":"external"},{"url":"https://huggingface.co/mistralai/Magistral-Small-2509","path":null,"label":"mistralai/Magistral-Small-2509","type":"external"},{"url":"https://huggingface.co/mistralai/Magistral-Small-2506","path":null,"label":"mistralai/Magistral-Small-2506","type":"external"},{"url":"https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512","path":null,"label":"mistralai/Ministral-3-3B-Reasoning-2512","type":"external"},{"url":"https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16","path":null,"label":"mistralai/Ministral-3-8B-Instruct-2512-BF16","type":"external"},{"url":"https://huggingface.co/mistralai/Devstral-Small-2507","path":null,"label":"mistralai/Devstral-Small-2507","type":"external"},{"url":"https://github.com/mistralai/mistral-inference","path":null,"label":"mistralai/mistral-inference","type":"external"},{"url":"https://github.com/mistralai/mistral-vibe","path":null,"label":"mistralai/mistral-vibe","type":"external"},{"url":"https://github.com/mistralai/mistral-vibe/releases/tag/v2.14.0","path":null,"label":"mistralai/mistral-vibe","type":"external"},{"url":"https://github.com/mistralai/mistral-finetune","path":null,"label":"mistralai/mistral-finetune","type":"external"},{"url":"https://github.com/mistralai/cookbook","path":null,"label":"mistralai/cookbook","type":"external"},{"url":"https://github.com/mistralai/mistral-common","path":null,"label":"mistralai/mistral-common","type":"external"},{"url":"https://github.com/mistralai/mistral-common/releases/tag/v1.11.3","path":null,"label":"mistralai/mistral-common","type":"external"},{"url":"https://github.com/mistralai/client-python","path":null,"label":"mistralai/client-python","type":"external"},{"url":"https://github.com/mistralai/client-python/releases/tag/v2.4.9","path":null,"label":"mistralai/client-python","type":"external"},{"url":"https://github.com/mistralai/client-ts","path":null,"label":"mistralai/client-ts","type":"external"},{"url":"https://github.com/mistralai/client-ts/releases/tag/packages/mistralai-gcp/v2.0.0","path":null,"label":"mistralai/client-ts","type":"external"},{"url":"https://github.com/mistralai/client-ts/releases/tag/packages/mistralai-azure/v2.0.0","path":null,"label":"mistralai/client-ts","type":"external"},{"url":"https://mistral.ai","path":null,"label":"mistral.ai","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"minimax","url":"https://onlylabs.fyi/analysis/minimax","json_url":"https://onlylabs.fyi/analysis/minimax/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/minimax/evidence.json","dossier_url":"https://onlylabs.fyi/labs/minimax","org":{"slug":"minimax","name":"MiniMax","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.minimax.io/"},"title":"MiniMax analysis","summary":"MiniMax is shipping fast-iterating, open-weight large language models — the MiniMax-M2 family dominates its current footprint, with the latest MiniMax-M2.7 (~229B params) already pulling 2.5M 30-day downloads on Hugging Face. Alongside the models it is building out the surrounding agent tooling (a CLI, an MCP server, an agent harness, and a published \"skills\" library), signaling a play to own not just the weights…","markdown":"## Thesis\n\nMiniMax is shipping fast-iterating, open-weight large language models — the MiniMax-M2 family dominates its current footprint, with the latest **MiniMax-M2.7** (~229B params) already pulling 2.5M 30-day downloads on Hugging Face. Alongside the models it is building out the surrounding agent tooling (a CLI, an MCP server, an agent harness, and a published \"skills\" library), signaling a play to own not just the weights but the developer surface around them.\n\n## Shipping\n\n- **MiniMax-M2.7** ([HF](https://huggingface.co/MiniMaxAI/MiniMax-M2.7), ~229B params) — its most-downloaded model at **2,498,939** 30-day downloads, 1,193 likes. The matching [MiniMax-M2.7 repo](https://github.com/MiniMax-AI/MiniMax-M2.7) has 334 stars.\n- **MiniMax-M2.5** ([HF](https://huggingface.co/MiniMaxAI/MiniMax-M2.5), ~229B params) — **678,819** downloads, 1,493 likes; [repo](https://github.com/MiniMax-AI/MiniMax-M2.5) at 584 stars.\n- **MiniMax-VL-01** ([HF](https://huggingface.co/MiniMaxAI/MiniMax-VL-01), ~456B params) — its vision-language model, **187,339** downloads, 285 likes.\n- **MiniMax-M2** ([HF](https://huggingface.co/MiniMaxAI/MiniMax-M2)) — **139,172** downloads and 1,495 likes (the most-liked card in the set); [MiniMax-M2 repo](https://github.com/MiniMax-AI/MiniMax-M2) at 2,595 stars.\n- **MiniMax-M1** (40k and 80k context variants) — [MiniMax-M1-40k](https://huggingface.co/MiniMaxAI/MiniMax-M1-40k) (~456B params) at 59,020 downloads; [MiniMax-M1-80k](https://huggingface.co/MiniMaxAI/MiniMax-M1-80k) at 846 downloads but 692 likes. Code in the [MiniMax-M1 repo](https://github.com/MiniMax-AI/MiniMax-M1) (3,153 stars).\n- **MiniMax-Text-01** ([HF](https://huggingface.co/MiniMaxAI/MiniMax-Text-01), ~456B params) and the [MiniMax-01 repo](https://github.com/MiniMax-AI/MiniMax-01) (3,426 stars) anchor the earlier text generation.\n- **SynLogic** reasoning models — [SynLogic-32B](https://huggingface.co/MiniMaxAI/SynLogic-32B) and [SynLogic-7B](https://huggingface.co/MiniMaxAI/SynLogic-7B) — and **VTP** ([VTP-Large-f16d64](https://huggingface.co/MiniMaxAI/VTP-Large-f16d64), [VTP repo](https://github.com/MiniMax-AI/VTP) at 490 stars) round out smaller, more specialized releases.\n- **Tooling/agent surface:** the [MiniMax CLI](https://github.com/MiniMax-AI/cli) (1,893 stars) is on an active release cadence — latest tag [v1.0.16](https://github.com/MiniMax-AI/cli/releases/tag/v1.0.16), with v1.0.12–v1.0.15 preceding it. Supporting repos include [MiniMax-MCP](https://github.com/MiniMax-AI/MiniMax-MCP) (1,506 stars), [Mini-Agent](https://github.com/MiniMax-AI/Mini-Agent) (2,759 stars), and [OpenRoom](https://github.com/MiniMax-AI/OpenRoom) (1,205 stars).\n\n## Research themes\n\nNo first-party writing captured yet.\n\n## Hiring & scaling\n\nThe open roles read as a broad, early-career talent pipeline rather than targeted senior hires: a **Top Talent Program**, **Graduate Recruitment 2026**, **Intern Recruitment 2027**, **Regular Internship**, and **Social Recruitment** (no locations captured). The mix — multiple intern/graduate tracks plus a \"Top Talent\" flagship — signals investment in building bench depth and recruiting ahead of need, consistent with a fast-shipping lab scaling its headcount rather than filling a few specialist gaps.\n\n## Traction highlights\n\n- **Most-downloaded model:** [MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) at 2,498,939 30-day downloads, far ahead of [MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) at 678,819.\n- **Most-starred repo:** [MiniMax-AI/skills](https://github.com/MiniMax-AI/skills) with 12,468 stars — its single biggest GitHub draw, outpacing every model repo. Next are [MiniMax-01](https://github.com/MiniMax-AI/MiniMax-01) (3,426) and [MiniMax-M1](https://github.com/MiniMax-AI/MiniMax-M1) (3,153).\n- **Hacker News:** thin traction — the only captured thread, [MiniMax-AI/MiniMax-M2](https://github.com/MiniMax-AI/MiniMax-M2), drew just 2 points and 0 comments.\n\n## Sources\n\n- [MiniMax homepage](https://www.minimax.io/)\n- [MiniMax-M2.7 (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)\n- [MiniMax-M2.5 (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)\n- [MiniMax-VL-01 (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-VL-01)\n- [MiniMax-M2 (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-M2)\n- [MiniMax-M1-40k (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-M1-40k)\n- [MiniMax-Text-01 (Hugging Face)](https://huggingface.co/MiniMaxAI/MiniMax-Text-01)\n- [SynLogic-32B (Hugging Face)](https://huggingface.co/MiniMaxAI/SynLogic-32B)\n- [MiniMax-AI/skills (GitHub)](https://github.com/MiniMax-AI/skills)\n- [MiniMax-AI/MiniMax-M2 (GitHub)](https://github.com/MiniMax-AI/MiniMax-M2)\n- [MiniMax-AI/cli (GitHub)](https://github.com/MiniMax-AI/cli) — [v1.0.16 release](https://github.com/MiniMax-AI/cli/releases/tag/v1.0.16)\n- [MiniMax-AI/MiniMax-MCP (GitHub)](https://github.com/MiniMax-AI/MiniMax-MCP)\n- [MiniMax-AI/Mini-Agent (GitHub)](https://github.com/MiniMax-AI/Mini-Agent)","generated_at":"2026-06-08T15:59:09.182+00:00","citations":[{"url":"https://huggingface.co/MiniMaxAI/MiniMax-M2.7","path":null,"label":"MiniMaxAI/MiniMax-M2.7","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-M2.7","path":null,"label":"MiniMax-AI/MiniMax-M2.7","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-M2.5","path":null,"label":"MiniMaxAI/MiniMax-M2.5","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-M2.5","path":null,"label":"MiniMax-AI/MiniMax-M2.5","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-VL-01","path":null,"label":"MiniMaxAI/MiniMax-VL-01","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-M2","path":null,"label":"MiniMaxAI/MiniMax-M2","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-M2","path":null,"label":"MiniMax-AI/MiniMax-M2","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-M1-40k","path":null,"label":"MiniMaxAI/MiniMax-M1-40k","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-M1-80k","path":null,"label":"MiniMaxAI/MiniMax-M1-80k","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-M1","path":null,"label":"MiniMax-AI/MiniMax-M1","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/MiniMax-Text-01","path":null,"label":"MiniMaxAI/MiniMax-Text-01","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-01","path":null,"label":"MiniMax-AI/MiniMax-01","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/SynLogic-32B","path":null,"label":"MiniMaxAI/SynLogic-32B","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/SynLogic-7B","path":null,"label":"MiniMaxAI/SynLogic-7B","type":"external"},{"url":"https://huggingface.co/MiniMaxAI/VTP-Large-f16d64","path":null,"label":"MiniMaxAI/VTP-Large-f16d64","type":"external"},{"url":"https://github.com/MiniMax-AI/VTP","path":null,"label":"MiniMax-AI/VTP","type":"external"},{"url":"https://github.com/MiniMax-AI/cli","path":null,"label":"MiniMax-AI/cli","type":"external"},{"url":"https://github.com/MiniMax-AI/cli/releases/tag/v1.0.16","path":null,"label":"MiniMax-AI/cli","type":"external"},{"url":"https://github.com/MiniMax-AI/MiniMax-MCP","path":null,"label":"MiniMax-AI/MiniMax-MCP","type":"external"},{"url":"https://github.com/MiniMax-AI/Mini-Agent","path":null,"label":"MiniMax-AI/Mini-Agent","type":"external"},{"url":"https://github.com/MiniMax-AI/OpenRoom","path":null,"label":"MiniMax-AI/OpenRoom","type":"external"},{"url":"https://github.com/MiniMax-AI/skills","path":null,"label":"MiniMax-AI/skills","type":"external"},{"url":"https://www.minimax.io/","path":null,"label":"minimax.io","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"microsoft","url":"https://onlylabs.fyi/analysis/microsoft","json_url":"https://onlylabs.fyi/analysis/microsoft/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/microsoft/evidence.json","dossier_url":"https://onlylabs.fyi/labs/microsoft","org":{"slug":"microsoft","name":"Microsoft","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.microsoft.com"},"title":"Microsoft analysis","summary":"Microsoft's public footprint is dominated by its developer-tools and platform empire — vscode (186k stars), PowerToys, TypeScript, terminal, and playwright — but the AI-specific signal sits in Microsoft Research, which is releasing small, domain-specific foundation models (materials, the electric grid) on Hugging Face rather than chasing a single frontier LLM. Recent shipping concentrates on agent infrastructure (an…","markdown":"## Thesis\nMicrosoft's public footprint is dominated by its developer-tools and platform empire — [vscode](https://github.com/microsoft/vscode) (186k stars), [PowerToys](https://github.com/microsoft/PowerToys), [TypeScript](https://github.com/microsoft/TypeScript), [terminal](https://github.com/microsoft/terminal), and [playwright](https://github.com/microsoft/playwright) — but the AI-specific signal sits in Microsoft Research, which is releasing small, domain-specific foundation models (materials, the electric grid) on Hugging Face rather than chasing a single frontier LLM. Recent shipping concentrates on agent infrastructure (an [agent-host-protocol](https://github.com/microsoft/agent-host-protocol/releases/tag/v0.3.0) spec, multiple \"solution-accelerator\" repos) and applied science (MatterSim, Dayhoff, GridSFM, Skala).\n\n## Shipping\n**Hugging Face models.** The most-downloaded release is [microsoft/skala-1.1](https://huggingface.co/microsoft/skala-1.1) at **163,473** downloads, with [microsoft/skala-1.0](https://huggingface.co/microsoft/skala-1.0) adding **38,616** (plus [skala-baselines](https://huggingface.co/microsoft/skala-baselines), 131). The \"Lens\" family draws the most likes: [microsoft/Lens](https://huggingface.co/microsoft/Lens) (3,882 downloads, 166 likes), [microsoft/Lens-Turbo](https://huggingface.co/microsoft/Lens-Turbo) (2,390 / 139 likes), and [microsoft/Lens-Base](https://huggingface.co/microsoft/Lens-Base) (1,176). [microsoft/Phi-Ground-Any](https://huggingface.co/microsoft/Phi-Ground-Any) (1,447) extends the Phi line. A large set of checkpoints ships under the **Dayhoff** family — both 170M-parameter ([Dayhoff-170M-GRS-SS-134000](https://huggingface.co/microsoft/Dayhoff-170M-GRS-SS-134000), 148 downloads) and 3B-parameter ([Dayhoff-3b-UR90-10000](https://huggingface.co/microsoft/Dayhoff-3b-UR90-10000), 76) variants — each in low-double/triple-digit downloads, consistent with a research-stage model line.\n\n**GitHub releases.** Recent tags skew toward agent and data-platform accelerators: [agent-host-protocol v0.3.0](https://github.com/microsoft/agent-host-protocol/releases/tag/v0.3.0) (and a paired [spec/v0.3.0](https://github.com/microsoft/agent-host-protocol/releases/tag/spec/v0.3.0)), [agentic-applications-for-unified-data-foundation-solution-accelerator v1.25.4](https://github.com/microsoft/agentic-applications-for-unified-data-foundation-solution-accelerator/releases/tag/v1.25.4), [Document-Knowledge-Mining-Solution-Accelerator v1.9.0](https://github.com/microsoft/Document-Knowledge-Mining-Solution-Accelerator/releases/tag/v1.9.0), [Conversation-Knowledge-Mining-Solution-Accelerator v3.23.2](https://github.com/microsoft/Conversation-Knowledge-Mining-Solution-Accelerator/releases/tag/v3.23.2), and [content-generation-solution-accelerator v2.7.0](https://github.com/microsoft/content-generation-solution-accelerator/releases/tag/v2.7.0). Platform/infra releases also landed: [WSL 2.7.8](https://github.com/microsoft/WSL/releases/tag/2.7.8), [component-detection v7.1.3](https://github.com/microsoft/component-detection/releases/tag/v7.1.3), and [openvmm-deps 0.3.0-48](https://github.com/microsoft/openvmm-deps/releases/tag/0.3.0-48).\n\n**Top repos.** Beyond the flagship editors and tools above, the AI-relevant standouts are [microsoft/graphrag](https://github.com/microsoft/graphrag) (33,483 stars), [microsoft/semantic-kernel](https://github.com/microsoft/semantic-kernel) (28,061), [microsoft/onnxruntime](https://github.com/microsoft/onnxruntime) (20,764), and [microsoft/typescript-go](https://github.com/microsoft/typescript-go) (25,618).\n\n## Research themes\nMicrosoft Research's recent first-party writing clusters around AI-for-science and agent reliability:\n- **AI for the physical world / energy:** [GridSFM: A new, small foundation model for the electric grid](https://www.microsoft.com/en-us/research/blog/gridsfm-a-new-small-foundation-model-for-the-electric-grid/) and [Building realistic electric transmission grid dataset at scale](https://www.microsoft.com/en-us/research/blog/building-realistic-electric-transmission-grid-dataset-at-scale-a-pipeline-from-open-dataset/).\n- **AI for materials:** [Advancing AI for materials with MatterSim: experimental synthesis, faster simulation, and multi-task models](https://www.microsoft.com/en-us/research/blog/advancing-ai-for-materials-with-mattersim-experimental-synthesis-faster-simulation-and-multi-task-models/).\n- **Agent reliability and human-AI augmentation:** [Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability](https://www.microsoft.com/en-us/research/blog/further-notes-on-our-recent-research-on-ai-delegation-and-long-horizon-reliability/) and [Extending Human Intelligence Through AI](https://www.microsoft.com/en-us/research/blog/extending-human-intelligence-through-ai/).\n- **Identity / trust and applied data tooling:** [Vega: Zero-knowledge proofs for digital identity in the age of AI](https://www.microsoft.com/en-us/research/blog/vega-zero-knowledge-proofs-for-digital-identity-in-the-age-of-ai/) and [Data Formulator 0.7: AI-powered data analytics for enterprise data](https://www.microsoft.com/en-us/research/blog/data-formulator-0-7-ai-powered-data-analytics-for-enterprise-data/).\n- **Systems:** [mimalloc: A new, high-performance, scalable memory allocator for the modern era](https://www.microsoft.com/en-us/research/blog/mimalloc-a-high-performance-scalable-memory-allocator-for-the-modern-era/).\n\n## Hiring & scaling\nNo careers data captured yet.\n\n## Traction highlights\nOn Hacker News, Microsoft's Rust and systems work draws the most attention: [microsoft/windows-rs](https://github.com/microsoft/windows-rs) leads at **785 points / 293 comments**, followed by [microsoft/playwright](https://github.com/microsoft/playwright) (383 / 140), [microsoft/garnet](https://github.com/microsoft/garnet) (370 / 118), [microsoft/windows-drivers-rs](https://github.com/microsoft/windows-drivers-rs) (339 / 142), [microsoft/ebpf-for-windows](https://github.com/microsoft/ebpf-for-windows) (294 / 165), and [microsoft/coreutils](https://github.com/microsoft/coreutils) (225 points but a high 246 comments). By stars, the most-followed repos are [vscode](https://github.com/microsoft/vscode) (186,063), [PowerToys](https://github.com/microsoft/PowerToys) (133,906), and [TypeScript](https://github.com/microsoft/TypeScript) (109,159). The most-downloaded model is [microsoft/skala-1.1](https://huggingface.co/microsoft/skala-1.1) (163,473 downloads); the most-liked is [microsoft/Lens](https://huggingface.co/microsoft/Lens) (166 likes).\n\n## Sources\n- https://huggingface.co/microsoft/skala-1.1\n- https://huggingface.co/microsoft/skala-1.0\n- https://huggingface.co/microsoft/Lens\n- https://huggingface.co/microsoft/Lens-Turbo\n- https://huggingface.co/microsoft/Phi-Ground-Any\n- https://huggingface.co/microsoft/Dayhoff-3b-UR90-10000\n- https://github.com/microsoft/vscode\n- https://github.com/microsoft/graphrag\n- https://github.com/microsoft/semantic-kernel\n- https://github.com/microsoft/typescript-go\n- https://github.com/microsoft/agent-host-protocol/releases/tag/v0.3.0\n- https://github.com/microsoft/agentic-applications-for-unified-data-foundation-solution-accelerator/releases/tag/v1.25.4\n- https://github.com/microsoft/WSL/releases/tag/2.7.8\n- https://github.com/microsoft/windows-rs\n- https://github.com/microsoft/garnet\n- https://github.com/microsoft/coreutils\n- https://www.microsoft.com/en-us/research/blog/gridsfm-a-new-small-foundation-model-for-the-electric-grid/\n- https://www.microsoft.com/en-us/research/blog/advancing-ai-for-materials-with-mattersim-experimental-synthesis-faster-simulation-and-multi-task-models/\n- https://www.microsoft.com/en-us/research/blog/further-notes-on-our-recent-research-on-ai-delegation-and-long-horizon-reliability/\n- https://www.microsoft.com/en-us/research/blog/vega-zero-knowledge-proofs-for-digital-identity-in-the-age-of-ai/\n- https://www.microsoft.com/en-us/research/blog/data-formulator-0-7-ai-powered-data-analytics-for-enterprise-data/","generated_at":"2026-06-08T15:59:09.069+00:00","citations":[{"url":"https://github.com/microsoft/vscode","path":null,"label":"microsoft/vscode","type":"external"},{"url":"https://github.com/microsoft/PowerToys","path":null,"label":"microsoft/PowerToys","type":"external"},{"url":"https://github.com/microsoft/TypeScript","path":null,"label":"microsoft/TypeScript","type":"external"},{"url":"https://github.com/microsoft/terminal","path":null,"label":"microsoft/terminal","type":"external"},{"url":"https://github.com/microsoft/playwright","path":null,"label":"microsoft/playwright","type":"external"},{"url":"https://github.com/microsoft/agent-host-protocol/releases/tag/v0.3.0","path":null,"label":"microsoft/agent-host-protocol","type":"external"},{"url":"https://huggingface.co/microsoft/skala-1.1","path":null,"label":"microsoft/skala-1.1","type":"external"},{"url":"https://huggingface.co/microsoft/skala-1.0","path":null,"label":"microsoft/skala-1.0","type":"external"},{"url":"https://huggingface.co/microsoft/skala-baselines","path":null,"label":"microsoft/skala-baselines","type":"external"},{"url":"https://huggingface.co/microsoft/Lens","path":null,"label":"microsoft/Lens","type":"external"},{"url":"https://huggingface.co/microsoft/Lens-Turbo","path":null,"label":"microsoft/Lens-Turbo","type":"external"},{"url":"https://huggingface.co/microsoft/Lens-Base","path":null,"label":"microsoft/Lens-Base","type":"external"},{"url":"https://huggingface.co/microsoft/Phi-Ground-Any","path":null,"label":"microsoft/Phi-Ground-Any","type":"external"},{"url":"https://huggingface.co/microsoft/Dayhoff-170M-GRS-SS-134000","path":null,"label":"microsoft/Dayhoff-170M-GRS-SS-134000","type":"external"},{"url":"https://huggingface.co/microsoft/Dayhoff-3b-UR90-10000","path":null,"label":"microsoft/Dayhoff-3b-UR90-10000","type":"external"},{"url":"https://github.com/microsoft/agent-host-protocol/releases/tag/spec/v0.3.0","path":null,"label":"microsoft/agent-host-protocol","type":"external"},{"url":"https://github.com/microsoft/agentic-applications-for-unified-data-foundation-solution-accelerator/releases/tag/v1.25.4","path":null,"label":"microsoft/agentic-applications-for-unified-data-foundation-solution-accelerator","type":"external"},{"url":"https://github.com/microsoft/Document-Knowledge-Mining-Solution-Accelerator/releases/tag/v1.9.0","path":null,"label":"microsoft/Document-Knowledge-Mining-Solution-Accelerator","type":"external"},{"url":"https://github.com/microsoft/Conversation-Knowledge-Mining-Solution-Accelerator/releases/tag/v3.23.2","path":null,"label":"microsoft/Conversation-Knowledge-Mining-Solution-Accelerator","type":"external"},{"url":"https://github.com/microsoft/content-generation-solution-accelerator/releases/tag/v2.7.0","path":null,"label":"microsoft/content-generation-solution-accelerator","type":"external"},{"url":"https://github.com/microsoft/WSL/releases/tag/2.7.8","path":null,"label":"microsoft/WSL","type":"external"},{"url":"https://github.com/microsoft/component-detection/releases/tag/v7.1.3","path":null,"label":"microsoft/component-detection","type":"external"},{"url":"https://github.com/microsoft/openvmm-deps/releases/tag/0.3.0-48","path":null,"label":"microsoft/openvmm-deps","type":"external"},{"url":"https://github.com/microsoft/graphrag","path":null,"label":"microsoft/graphrag","type":"external"},{"url":"https://github.com/microsoft/semantic-kernel","path":null,"label":"microsoft/semantic-kernel","type":"external"},{"url":"https://github.com/microsoft/onnxruntime","path":null,"label":"microsoft/onnxruntime","type":"external"},{"url":"https://github.com/microsoft/typescript-go","path":null,"label":"microsoft/typescript-go","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/gridsfm-a-new-small-foundation-model-for-the-electric-grid/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/building-realistic-electric-transmission-grid-dataset-at-scale-a-pipeline-from-open-dataset/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/advancing-ai-for-materials-with-mattersim-experimental-synthesis-faster-simulation-and-multi-task-models/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/further-notes-on-our-recent-research-on-ai-delegation-and-long-horizon-reliability/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/extending-human-intelligence-through-ai/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/vega-zero-knowledge-proofs-for-digital-identity-in-the-age-of-ai/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/data-formulator-0-7-ai-powered-data-analytics-for-enterprise-data/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://www.microsoft.com/en-us/research/blog/mimalloc-a-high-performance-scalable-memory-allocator-for-the-modern-era/","path":null,"label":"microsoft.com/en-us","type":"external"},{"url":"https://github.com/microsoft/windows-rs","path":null,"label":"microsoft/windows-rs","type":"external"},{"url":"https://github.com/microsoft/garnet","path":null,"label":"microsoft/garnet","type":"external"},{"url":"https://github.com/microsoft/windows-drivers-rs","path":null,"label":"microsoft/windows-drivers-rs","type":"external"},{"url":"https://github.com/microsoft/ebpf-for-windows","path":null,"label":"microsoft/ebpf-for-windows","type":"external"},{"url":"https://github.com/microsoft/coreutils","path":null,"label":"microsoft/coreutils","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"meta-ai","url":"https://onlylabs.fyi/analysis/meta-ai","json_url":"https://onlylabs.fyi/analysis/meta-ai/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/meta-ai/evidence.json","dossier_url":"https://onlylabs.fyi/labs/meta-ai","org":{"slug":"meta-ai","name":"Meta AI (Llama)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://ai.meta.com"},"title":"Meta AI (Llama) analysis","summary":"Meta AI is the open-weight anchor of the frontier-model field: it ships the Llama family under permissive licenses and lets the ecosystem do distribution, while pivoting its newest generation (Llama 4) to mixture-of-experts. Alongside the models it is building out the surrounding tooling — a hosted Llama API (Python/TypeScript SDKs), the PurpleLlama/Llama-Guard safety stack, and developer cookbooks — and its public…","markdown":"## Thesis\n\nMeta AI is the open-weight anchor of the frontier-model field: it ships the Llama family under permissive licenses and lets the ecosystem do distribution, while pivoting its newest generation (Llama 4) to mixture-of-experts. Alongside the models it is building out the surrounding tooling — a hosted Llama API (Python/TypeScript SDKs), the PurpleLlama/Llama-Guard safety stack, and developer cookbooks — and its public engineering writing is dominated by AI infrastructure and applied-LLM systems work rather than model announcements.\n\n## Shipping\n\nThe footprint is led by the Llama checkpoints on Hugging Face. The most-pulled by far is [`meta-llama/Llama-3.1-8B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) at **11,216,853** 30-day downloads (6,013 likes), followed by the small Llama 3.2 line — [`Llama-3.2-1B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) at **8,117,344**, [`Llama-3.2-1B`](https://huggingface.co/meta-llama/Llama-3.2-1B) at **2,338,719**, and [`Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) at **1,693,307**. The flagship dense model [`Llama-3.3-70B-Instruct`](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) draws **787,281** downloads (2,805 likes), and the 405B [`Llama-3.1-405B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct) sits at **219,986**.\n\nThe newest generation is MoE: [`Llama-4-Scout-17B-16E-Instruct`](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) (108B total params, 16 experts) at **452,362** downloads and [`Llama-4-Maverick-17B-128E-Instruct`](https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct) (401B total, 128 experts) at **33,079**. Multimodal shows up via [`Llama-3.2-11B-Vision-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct) (**173,277**). A notable share of the catalog is safety tooling: [`Prompt-Guard-86M`](https://huggingface.co/meta-llama/Prompt-Guard-86M) (**697,663**), [`Llama-Guard-4-12B`](https://huggingface.co/meta-llama/Llama-Guard-4-12B) (**152,961**), [`Llama-Prompt-Guard-2-86M`](https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M) (**136,048**), plus the [`Llama-Guard-3-8B`](https://huggingface.co/meta-llama/Llama-Guard-3-8B) and [`Llama-Guard-3-1B`](https://huggingface.co/meta-llama/Llama-Guard-3-1B) classifiers.\n\nOn GitHub the legacy [`meta-llama/llama`](https://github.com/meta-llama/llama) repo still leads at **59,454** stars, with [`llama3`](https://github.com/meta-llama/llama3) at 29,287, [`llama-cookbook`](https://github.com/meta-llama/llama-cookbook) at 18,346, [`codellama`](https://github.com/meta-llama/codellama) at 16,314, [`llama-models`](https://github.com/meta-llama/llama-models) at 7,625, and the safety repo [`PurpleLlama`](https://github.com/meta-llama/PurpleLlama) at 4,210. Recent release activity is concentrated on the hosted API surface: [`llama-api-python v0.6.0`](https://github.com/meta-llama/llama-api-python/releases/tag/v0.6.0) and [`llama-api-typescript v0.3.0`](https://github.com/meta-llama/llama-api-typescript/releases/tag/v0.3.0) are the latest of a steady cadence of SDK point releases, alongside [`llama-verifications`](https://github.com/meta-llama/llama-verifications/releases/tag/v0.1.20.1.2rc2). Newer data/ops repos — [`synthetic-data-kit`](https://github.com/meta-llama/synthetic-data-kit) (1,597 stars) and [`prompt-ops`](https://github.com/meta-llama/prompt-ops) (820) — round out the developer-tooling push.\n\n## Research themes\n\nMeta's captured engineering writing skews toward AI *infrastructure and applied LLM systems* over model releases:\n\n- **LLM inference and GPU systems at scale** — [\"Scaling LLM Inference: Innovations in Tensor Parallelism, Context Parallelism, and Expert Parallelism\"](https://engineering.fb.com/2025/10/17/ai-research/scaling-llm-inference-innovations-tensor-parallelism-context-parallelism-expert-parallelism/) (tied to the Meta AI App), [\"RCCLX: Innovating GPU Communications on AMD Platforms\"](https://engineering.fb.com/2026/02/24/data-center-engineering/rrcclx-innovating-gpu-communications-amd-platforms-meta/) (open-sourced, AMD/Torchcomms), and [\"Meta's Infrastructure Evolution and the Advent of AI\"](https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/).\n- **Vector search** — [\"Accelerating GPU indexes in Faiss with NVIDIA cuVS\"](https://engineering.fb.com/2025/05/08/data-infrastructure/accelerating-gpu-indexes-in-faiss-with-nvidia-cuvs/), reporting up to 4.7x faster IVF build and 8.1x lower search latency in Faiss v1.10.\n- **LLMs applied to software engineering** — [\"Diff Risk Score: AI-driven risk-aware software development\"](https://engineering.fb.com/2025/08/06/developer-tools/diff-risk-score-drs-ai-risk-aware-software-development-meta/) (a fine-tuned Llama predicting production-incident risk) and [\"LLMs Are the Key to Mutation Testing and Better Compliance\"](https://engineering.fb.com/2025/09/30/security/llms-are-the-key-to-mutation-testing-and-better-compliance/) (the ACH compliance-hardening tool).\n- **AI for science and the physical/AR-VR world** — [\"Using AI to make lower-carbon, faster-curing concrete\"](https://engineering.fb.com/2025/07/16/data-center-engineering/ai-make-lower-carbon-faster-curing-concrete/) (Bayesian optimization via BoTorch/Ax), [\"Meta 3D AssetGen: Generating 3D Worlds With AI\"](https://engineering.fb.com/2025/09/29/virtual-reality/assetgen-generating-3d-worlds-with-ai/), and [\"Building a human-computer interface for everyone\"](https://engineering.fb.com/2025/08/04/virtual-reality/building-a-human-computer-interface-for-everyone-meta-tech-podcast/) (Reality Labs sEMG wristband).\n\n## Hiring & scaling\n\nThe 15 captured roles read as broad product-and-platform scaling rather than a pure research build-out. Engineering is the largest bucket — multiple **Software Engineer** openings including Product, **Infrastructure**, and **AR/VR** (Redmond, WA), plus a **Machine Learning Engineer** (Palo Alto) — with research demand showing in two **Research Scientist, AI** posts (New York and Palo Alto). Supporting functions span the full product org: **Data Scientist / Data Scientist, Analytics** (Menlo Park and New York), **Technical Program Manager** (Seattle), **Product Manager / Product Designer / Product Marketing Manager**, and a **Security Engineer** (Menlo Park). Geographically the center of gravity is Menlo Park, with secondary clusters in New York, the Bay Area, and the Seattle/Redmond corridor — consistent with both the AI App / infrastructure work and the Reality Labs AR/VR investment surfaced in the writing.\n\n## Traction highlights\n\n- **Most-downloaded model:** [`Llama-3.1-8B-Instruct`](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) at **11.2M** 30-day downloads, with the Llama 3.2 1B/3B small models close behind (8.1M / 1.7M+).\n- **Most-starred repo:** [`meta-llama/llama`](https://github.com/meta-llama/llama) at **59,454** stars, followed by [`llama3`](https://github.com/meta-llama/llama3) (29,287) and [`llama-cookbook`](https://github.com/meta-llama/llama-cookbook) (18,346).\n- **Hacker News:** captured traction is thin — [\"Meta's Infrastructure Evolution and the Advent of AI\"](https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/) reached only **4 points / 0 comments** and [\"LLMs Are the Key to Mutation Testing and Better Compliance\"](https://engineering.fb.com/2025/09/30/security/llms-are-the-key-to-mutation-testing-and-better-compliance/) **2 points / 1 comment**. The distribution story is on Hugging Face and GitHub, not HN.\n\n## Sources\n\n- https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct\n- https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct\n- https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct\n- https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct\n- https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct\n- https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct\n- https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct\n- https://huggingface.co/meta-llama/Prompt-Guard-86M\n- https://huggingface.co/meta-llama/Llama-Guard-4-12B\n- https://github.com/meta-llama/llama\n- https://github.com/meta-llama/llama3\n- https://github.com/meta-llama/llama-cookbook\n- https://github.com/meta-llama/PurpleLlama\n- https://github.com/meta-llama/llama-api-python/releases/tag/v0.6.0\n- https://github.com/meta-llama/llama-api-typescript/releases/tag/v0.3.0\n- https://engineering.fb.com/2025/10/17/ai-research/scaling-llm-inference-innovations-tensor-parallelism-context-parallelism-expert-parallelism/\n- https://engineering.fb.com/2025/05/08/data-infrastructure/accelerating-gpu-indexes-in-faiss-with-nvidia-cuvs/\n- https://engineering.fb.com/2025/08/06/developer-tools/diff-risk-score-drs-ai-risk-aware-software-development-meta/\n- https://engineering.fb.com/2025/09/30/security/llms-are-the-key-to-mutation-testing-and-better-compliance/\n- https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/\n- https://engineering.fb.com/2025/09/29/virtual-reality/assetgen-generating-3d-worlds-with-ai/\n- https://engineering.fb.com/2026/02/24/data-center-engineering/rrcclx-innovating-gpu-communications-amd-platforms-meta/\n- https://ai.meta.com","generated_at":"2026-06-08T15:59:08.887+00:00","citations":[{"url":"https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct","path":null,"label":"meta-llama/Llama-3.1-8B-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct","path":null,"label":"meta-llama/Llama-3.2-1B-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.2-1B","path":null,"label":"meta-llama/Llama-3.2-1B","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct","path":null,"label":"meta-llama/Llama-3.2-3B-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct","path":null,"label":"meta-llama/Llama-3.3-70B-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct","path":null,"label":"meta-llama/Llama-3.1-405B-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct","path":null,"label":"meta-llama/Llama-4-Scout-17B-16E-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct","path":null,"label":"meta-llama/Llama-4-Maverick-17B-128E-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct","path":null,"label":"meta-llama/Llama-3.2-11B-Vision-Instruct","type":"external"},{"url":"https://huggingface.co/meta-llama/Prompt-Guard-86M","path":null,"label":"meta-llama/Prompt-Guard-86M","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-Guard-4-12B","path":null,"label":"meta-llama/Llama-Guard-4-12B","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M","path":null,"label":"meta-llama/Llama-Prompt-Guard-2-86M","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-Guard-3-8B","path":null,"label":"meta-llama/Llama-Guard-3-8B","type":"external"},{"url":"https://huggingface.co/meta-llama/Llama-Guard-3-1B","path":null,"label":"meta-llama/Llama-Guard-3-1B","type":"external"},{"url":"https://github.com/meta-llama/llama","path":null,"label":"meta-llama/llama","type":"external"},{"url":"https://github.com/meta-llama/llama3","path":null,"label":"meta-llama/llama3","type":"external"},{"url":"https://github.com/meta-llama/llama-cookbook","path":null,"label":"meta-llama/llama-cookbook","type":"external"},{"url":"https://github.com/meta-llama/codellama","path":null,"label":"meta-llama/codellama","type":"external"},{"url":"https://github.com/meta-llama/llama-models","path":null,"label":"meta-llama/llama-models","type":"external"},{"url":"https://github.com/meta-llama/PurpleLlama","path":null,"label":"meta-llama/PurpleLlama","type":"external"},{"url":"https://github.com/meta-llama/llama-api-python/releases/tag/v0.6.0","path":null,"label":"meta-llama/llama-api-python","type":"external"},{"url":"https://github.com/meta-llama/llama-api-typescript/releases/tag/v0.3.0","path":null,"label":"meta-llama/llama-api-typescript","type":"external"},{"url":"https://github.com/meta-llama/llama-verifications/releases/tag/v0.1.20.1.2rc2","path":null,"label":"meta-llama/llama-verifications","type":"external"},{"url":"https://github.com/meta-llama/synthetic-data-kit","path":null,"label":"meta-llama/synthetic-data-kit","type":"external"},{"url":"https://github.com/meta-llama/prompt-ops","path":null,"label":"meta-llama/prompt-ops","type":"external"},{"url":"https://engineering.fb.com/2025/10/17/ai-research/scaling-llm-inference-innovations-tensor-parallelism-context-parallelism-expert-parallelism/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2026/02/24/data-center-engineering/rrcclx-innovating-gpu-communications-amd-platforms-meta/","path":null,"label":"engineering.fb.com/2026","type":"external"},{"url":"https://engineering.fb.com/2025/09/29/data-infrastructure/metas-infrastructure-evolution-and-the-advent-of-ai/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/05/08/data-infrastructure/accelerating-gpu-indexes-in-faiss-with-nvidia-cuvs/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/08/06/developer-tools/diff-risk-score-drs-ai-risk-aware-software-development-meta/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/09/30/security/llms-are-the-key-to-mutation-testing-and-better-compliance/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/07/16/data-center-engineering/ai-make-lower-carbon-faster-curing-concrete/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/09/29/virtual-reality/assetgen-generating-3d-worlds-with-ai/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://engineering.fb.com/2025/08/04/virtual-reality/building-a-human-computer-interface-for-everyone-meta-tech-podcast/","path":null,"label":"engineering.fb.com/2025","type":"external"},{"url":"https://ai.meta.com","path":null,"label":"ai.meta.com","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"google-deepmind","url":"https://onlylabs.fyi/analysis/google-deepmind","json_url":"https://onlylabs.fyi/analysis/google-deepmind/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/google-deepmind/evidence.json","dossier_url":"https://onlylabs.fyi/labs/google-deepmind","org":{"slug":"google-deepmind","name":"Google (DeepMind / Gemini)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://deepmind.google/"},"title":"Google (DeepMind / Gemini) analysis","summary":"Google DeepMind is running a two-front strategy: shipping the proprietary Gemini 2.5 frontier line (Pro, Flash, Flash-Lite, Deep Think, Computer Use) as the consumer/enterprise product, while seeding an open-weights ecosystem around the Gemma family on Hugging Face. In parallel it keeps pushing \"AI for science\" — genomics (AlphaGenome), fusion plasma control, gravitational-wave instrumentation, and single-cell…","markdown":"## Thesis\n\nGoogle DeepMind is running a two-front strategy: shipping the proprietary Gemini 2.5 frontier line (Pro, Flash, Flash-Lite, Deep Think, Computer Use) as the consumer/enterprise product, while seeding an open-weights ecosystem around the Gemma family on Hugging Face. In parallel it keeps pushing \"AI for science\" — genomics (AlphaGenome), fusion plasma control, gravitational-wave instrumentation, and single-cell biology — as proof points that its models do real-world discovery, not just chat.\n\n## Shipping\n\nOn Hugging Face the open-weights footprint is dominated by the Gemma 4 generation. [`google/gemma-4-26B-A4B-it`](https://huggingface.co/google/gemma-4-26B-A4B-it) is the clear leader at **12,161,679** 30-day downloads (1,099 likes), an order of magnitude ahead of everything else. Behind it: [`google/gemma-4-31B`](https://huggingface.co/google/gemma-4-31B) (565,446), [`google/gemma-4-12B-it`](https://huggingface.co/google/gemma-4-12B-it) (434,969), [`google/gemma-4-31B-it-assistant`](https://huggingface.co/google/gemma-4-31B-it-assistant) (408,902), and [`google/gemma-4-26B-A4B-it-assistant`](https://huggingface.co/google/gemma-4-26B-A4B-it-assistant) (159,886). The tail shows the breadth of the open program: vision encoders [`google/tipsv2-b14`](https://huggingface.co/google/tipsv2-b14) (18,165) and its l14/g14 variants, plus generative audio in [`google/magenta-realtime-2`](https://huggingface.co/google/magenta-realtime-2) (13,338).\n\nOn GitHub the most-starred repos reflect a research-tooling legacy more than the current Gemini push: [`deepmind-research`](https://github.com/google-deepmind/deepmind-research) (14,998 stars), [`alphafold`](https://github.com/google-deepmind/alphafold) (14,647), [`mujoco`](https://github.com/google-deepmind/mujoco) (13,791), [`sonnet`](https://github.com/google-deepmind/sonnet) (9,920), [`alphafold3`](https://github.com/google-deepmind/alphafold3) (8,147), and [`gemma`](https://github.com/google-deepmind/gemma) (5,356). Recent release activity is concentrated in the physics-sim and agent-tooling stack: [`mujoco 3.9.0`](https://github.com/google-deepmind/mujoco/releases/tag/3.9.0) and [`mujoco_warp v3.9.0`](https://github.com/google-deepmind/mujoco_warp/releases/tag/v3.9.0), [`gemma v4.0.1`](https://github.com/google-deepmind/gemma/releases/tag/v4.0.1), [`onetwo v0.5.0`](https://github.com/google-deepmind/onetwo/releases/tag/v0.5.0), [`open_spiel v1.6.15`](https://github.com/google-deepmind/open_spiel/releases/tag/v1.6.15), and a [`science-skills`](https://github.com/google-deepmind/science-skills/releases/tag/v1.0.2) package now at v1.0.2.\n\n## Research themes\n\nFour themes recur across first-party writing:\n\n- **Gemini 2.5 as a \"thinking\" frontier line.** Repeated updates push reasoning and product surface area: GA of [Gemini 2.5 Pro/Flash plus Flash-Lite](https://deepmind.google/blog/were-expanding-our-gemini-25-family-of-models/), the [Deep Think reasoning mode](https://deepmind.google/blog/try-deep-think-in-the-gemini-app/), [native audio dialog/generation](https://deepmind.google/blog/advanced-audio-dialog-and-generation-with-gemini-25/), and the agent-oriented [Gemini 2.5 Computer Use model](https://deepmind.google/blog/introducing-the-gemini-25-computer-use-model/). The framing is a \"[universal AI assistant](https://deepmind.google/blog/our-vision-for-building-a-universal-ai-assistant/)\" extending Gemini toward a world model.\n- **Generative media.** [Veo 3 and Imagen 4 plus the Flow filmmaking tool](https://deepmind.google/blog/fuel-your-creativity-with-new-generative-media-models-and-tools/), [Veo 3.1](https://deepmind.google/blog/introducing-veo-31-and-advanced-creative-capabilities/), and the \"[Nano Banana](https://deepmind.google/blog/image-editing-in-gemini-just-got-a-major-upgrade/)\" image-editing upgrade.\n- **AI for science.** [AlphaGenome](https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/) (now published in Nature), the Gemma-based [C2S-Scale 27B single-cell model that surfaced a cancer-therapy pathway](https://deepmind.google/blog/how-a-gemma-model-helped-discover-a-new-potential-cancer-therapy-pathway/), [tropical cyclone prediction with the U.S. National Hurricane Center](https://deepmind.google/blog/how-were-supporting-better-tropical-cyclone-prediction-with-ai/), a [fusion partnership with Commonwealth Fusion Systems](https://deepmind.google/blog/bringing-ai-to-the-next-generation-of-fusion-energy/), and [Deep Loop Shaping for gravitational-wave observatories](https://deepmind.google/blog/using-ai-to-perceive-the-universe-in-greater-depth/).\n- **Robotics, safety, and provenance.** On-device and agentic robotics ([Gemini Robotics On-Device](https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/), [Gemini Robotics 1.5](https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/)); responsibility work including [SynthID Detector](https://deepmind.google/blog/synthid-detector--a-new-portal-to-help-identify-ai-generated-content/), [CodeMender](https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/), [Backstory](https://deepmind.google/blog/exploring-the-context-of-online-images-with-backstory/), the [VaultGemma differentially-private LLM](https://deepmind.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/), and a third iteration of the [Frontier Safety Framework](https://deepmind.google/blog/strengthening-our-frontier-safety-framework/). Competitive-eval milestones include Gemini 2.5 Deep Think reaching [gold-medal level at the ICPC World Finals](https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/) and the [Kaggle Game Arena](https://deepmind.google/blog/rethinking-how-we-measure-ai-intelligence/) eval platform.\n\n## Hiring & scaling\n\nThe 15 open roles point to investment in materials science and applied/agentic AI. Multiple materials/intelligence roles appear — \"Research Scientist, Material Intelligence\" (London), \"Research Engineer, Materials Science\" (Mountain View) — reinforcing the AI-for-science theme. A cluster of \"Antigravity\" roles (\"Technical Program Manager, Antigravity\" and \"...Antigravity (Modeling & Evals),\" both Mountain View) plus \"Technical Program Manager, Agents Innovation\" (London) and two \"Applied AI\" roles (\"Staff Research Engineer, Applied AI\" Singapore; \"Manager, Applied AI Engineering\" London) signal a build-out around agents, modeling/evals, and productization. Internationalization shows up directly in \"Research Scientist: Multilingual, Multicultural and Multimodal LLM\" (Tokyo), and embodied AI in \"Research Scientist, HRI Research to Enable Collaborative Humanoid Robots\" (New York City). Geographically the footprint spans Mountain View, London, Singapore, Tokyo, and NYC.\n\n## Traction highlights\n\nHacker News interest skews toward agents and embodied/scientific systems rather than the core Gemini model drops. Top threads: [AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields](https://deepmind.google/blog/alphaevolve-impact/) (327 points, 149 comments), [Reimagining the mouse pointer for the AI era](https://deepmind.google/blog/ai-pointer/) (252 points, 213 comments), [SIMA 2: An Agent that Plays, Reasons, and Learns With You in Virtual 3D Worlds](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/) (238 points), and [Gemini Robotics-ER 1.6](https://deepmind.google/blog/gemini-robotics-er-1-6/) (219 points). The [`mujoco`](https://github.com/google-deepmind/mujoco) repo also surfaced on HN (116 points). On raw distribution, the standout is [`google/gemma-4-26B-A4B-it`](https://huggingface.co/google/gemma-4-26B-A4B-it) at over 12.1M 30-day downloads; on GitHub stars, [`deepmind-research`](https://github.com/google-deepmind/deepmind-research) (14,998) and [`alphafold`](https://github.com/google-deepmind/alphafold) (14,647) lead.\n\n## Sources\n\n- [Hugging Face: google/gemma-4-26B-A4B-it (12.1M downloads)](https://huggingface.co/google/gemma-4-26B-A4B-it)\n- [Hugging Face: google/gemma-4-31B](https://huggingface.co/google/gemma-4-31B)\n- [Hugging Face: google/magenta-realtime-2](https://huggingface.co/google/magenta-realtime-2)\n- [GitHub: google-deepmind/deepmind-research (14,998 stars)](https://github.com/google-deepmind/deepmind-research)\n- [GitHub: google-deepmind/alphafold (14,647 stars)](https://github.com/google-deepmind/alphafold)\n- [GitHub: google-deepmind/mujoco (13,791 stars)](https://github.com/google-deepmind/mujoco)\n- [GitHub release: google-deepmind/gemma v4.0.1](https://github.com/google-deepmind/gemma/releases/tag/v4.0.1)\n- [Blog: Our vision for building a universal AI assistant](https://deepmind.google/blog/our-vision-for-building-a-universal-ai-assistant/)\n- [Blog: We're expanding our Gemini 2.5 family of models](https://deepmind.google/blog/were-expanding-our-gemini-25-family-of-models/)\n- [Blog: Introducing the Gemini 2.5 Computer Use model](https://deepmind.google/blog/introducing-the-gemini-25-computer-use-model/)\n- [Blog: AlphaGenome — AI for better understanding the genome](https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/)\n- [Blog: How a Gemma model helped discover a new potential cancer therapy pathway](https://deepmind.google/blog/how-a-gemma-model-helped-discover-a-new-potential-cancer-therapy-pathway/)\n- [Blog: VaultGemma — the world's most capable differentially private LLM](https://deepmind.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/)\n- [Blog: Gemini achieves gold-medal level at the ICPC World Finals](https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/)\n- [HN: AlphaEvolve — Gemini-powered coding agent (327 points)](https://deepmind.google/blog/alphaevolve-impact/)\n- [HN: Reimagining the mouse pointer for the AI era (252 points)](https://deepmind.google/blog/ai-pointer/)\n- [HN: SIMA 2 (238 points)](https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/)","generated_at":"2026-06-08T15:59:08.768+00:00","citations":[{"url":"https://huggingface.co/google/gemma-4-26B-A4B-it","path":null,"label":"google/gemma-4-26B-A4B-it","type":"external"},{"url":"https://huggingface.co/google/gemma-4-31B","path":null,"label":"google/gemma-4-31B","type":"external"},{"url":"https://huggingface.co/google/gemma-4-12B-it","path":null,"label":"google/gemma-4-12B-it","type":"external"},{"url":"https://huggingface.co/google/gemma-4-31B-it-assistant","path":null,"label":"google/gemma-4-31B-it-assistant","type":"external"},{"url":"https://huggingface.co/google/gemma-4-26B-A4B-it-assistant","path":null,"label":"google/gemma-4-26B-A4B-it-assistant","type":"external"},{"url":"https://huggingface.co/google/tipsv2-b14","path":null,"label":"google/tipsv2-b14","type":"external"},{"url":"https://huggingface.co/google/magenta-realtime-2","path":null,"label":"google/magenta-realtime-2","type":"external"},{"url":"https://github.com/google-deepmind/deepmind-research","path":null,"label":"google-deepmind/deepmind-research","type":"external"},{"url":"https://github.com/google-deepmind/alphafold","path":null,"label":"google-deepmind/alphafold","type":"external"},{"url":"https://github.com/google-deepmind/mujoco","path":null,"label":"google-deepmind/mujoco","type":"external"},{"url":"https://github.com/google-deepmind/sonnet","path":null,"label":"google-deepmind/sonnet","type":"external"},{"url":"https://github.com/google-deepmind/alphafold3","path":null,"label":"google-deepmind/alphafold3","type":"external"},{"url":"https://github.com/google-deepmind/gemma","path":null,"label":"google-deepmind/gemma","type":"external"},{"url":"https://github.com/google-deepmind/mujoco/releases/tag/3.9.0","path":null,"label":"google-deepmind/mujoco","type":"external"},{"url":"https://github.com/google-deepmind/mujoco_warp/releases/tag/v3.9.0","path":null,"label":"google-deepmind/mujoco_warp","type":"external"},{"url":"https://github.com/google-deepmind/gemma/releases/tag/v4.0.1","path":null,"label":"google-deepmind/gemma","type":"external"},{"url":"https://github.com/google-deepmind/onetwo/releases/tag/v0.5.0","path":null,"label":"google-deepmind/onetwo","type":"external"},{"url":"https://github.com/google-deepmind/open_spiel/releases/tag/v1.6.15","path":null,"label":"google-deepmind/open_spiel","type":"external"},{"url":"https://github.com/google-deepmind/science-skills/releases/tag/v1.0.2","path":null,"label":"google-deepmind/science-skills","type":"external"},{"url":"https://deepmind.google/blog/were-expanding-our-gemini-25-family-of-models/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/try-deep-think-in-the-gemini-app/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/advanced-audio-dialog-and-generation-with-gemini-25/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/introducing-the-gemini-25-computer-use-model/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/our-vision-for-building-a-universal-ai-assistant/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/fuel-your-creativity-with-new-generative-media-models-and-tools/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/introducing-veo-31-and-advanced-creative-capabilities/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/image-editing-in-gemini-just-got-a-major-upgrade/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/how-a-gemma-model-helped-discover-a-new-potential-cancer-therapy-pathway/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/how-were-supporting-better-tropical-cyclone-prediction-with-ai/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/bringing-ai-to-the-next-generation-of-fusion-energy/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/using-ai-to-perceive-the-universe-in-greater-depth/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/synthid-detector--a-new-portal-to-help-identify-ai-generated-content/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/exploring-the-context-of-online-images-with-backstory/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/strengthening-our-frontier-safety-framework/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/rethinking-how-we-measure-ai-intelligence/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/alphaevolve-impact/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/ai-pointer/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/","path":null,"label":"deepmind.google/blog","type":"external"},{"url":"https://deepmind.google/blog/gemini-robotics-er-1-6/","path":null,"label":"deepmind.google/blog","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"deepseek","url":"https://onlylabs.fyi/analysis/deepseek","json_url":"https://onlylabs.fyi/analysis/deepseek/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/deepseek/evidence.json","dossier_url":"https://onlylabs.fyi/labs/deepseek","org":{"slug":"deepseek","name":"DeepSeek","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://www.deepseek.com/"},"title":"DeepSeek analysis","summary":"DeepSeek is shipping open-weight models at scale while simultaneously open-sourcing systems infrastructure alongside them. Its public footprint is split between flagship reasoning/general models (the R1 and V3 lines, both ~684.5B-param checkpoints) and a steady drumbeat of systems/infrastructure releases (DeepGEMM, DeepEP, FlashMLA, 3FS) — a model-plus-infra strategy. Flagship weights dominate downloads on Hugging…","markdown":"## Thesis\n\nDeepSeek is shipping open-weight models at scale while simultaneously open-sourcing systems infrastructure alongside them. Its public footprint is split between flagship reasoning/general models (the [R1](https://huggingface.co/deepseek-ai/DeepSeek-R1) and [V3](https://huggingface.co/deepseek-ai/DeepSeek-V3) lines, both ~684.5B-param checkpoints) and a steady drumbeat of systems/infrastructure releases ([DeepGEMM](https://github.com/deepseek-ai/DeepGEMM), [DeepEP](https://github.com/deepseek-ai/DeepEP), [FlashMLA](https://github.com/deepseek-ai/FlashMLA), [3FS](https://github.com/deepseek-ai/3FS)) — a model-plus-infra strategy.\n\n## Shipping\n\nFlagship weights dominate downloads on Hugging Face. [`DeepSeek-R1-0528`](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) leads at **6.44M** 30-day downloads (684.5B params), with the original [`DeepSeek-R1`](https://huggingface.co/deepseek-ai/DeepSeek-R1) at **5.69M** downloads and the highest like count on the slate (13,376). The general-model line continues with [`DeepSeek-V3.2`](https://huggingface.co/deepseek-ai/DeepSeek-V3.2) at **4.12M** downloads (685.4B params) and the earlier [`DeepSeek-V3`](https://huggingface.co/deepseek-ai/DeepSeek-V3) at **1.07M**; an experimental [`DeepSeek-V3.2-Exp`](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp) is also published (166K downloads).\n\nA second product axis is OCR/vision: [`DeepSeek-OCR`](https://huggingface.co/deepseek-ai/DeepSeek-OCR) (**2.70M** downloads, 3.34B params) and its successor [`DeepSeek-OCR-2`](https://huggingface.co/deepseek-ai/DeepSeek-OCR-2) (**1.84M**), plus the smaller VLM [`deepseek-vl2-tiny`](https://huggingface.co/deepseek-ai/deepseek-vl2-tiny) (843K).\n\nThe R1 reasoning behavior is widely propagated through distillations into other base models: [`R1-Distill-Qwen-1.5B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B) (772K), [`R1-Distill-Qwen-32B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) (554K), [`R1-Distill-Qwen-14B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) (547K), [`R1-Distill-Qwen-7B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) (531K), [`R1-Distill-Llama-8B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) (486K), and [`R1-0528-Qwen3-8B`](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B) (291K).\n\nOn GitHub, the model repos anchor the top of the stars chart — [`DeepSeek-V3`](https://github.com/deepseek-ai/DeepSeek-V3) (**103.7k** stars) and [`DeepSeek-R1`](https://github.com/deepseek-ai/DeepSeek-R1) (**92.0k**) — followed by [`awesome-deepseek-integration`](https://github.com/deepseek-ai/awesome-deepseek-integration) (37.8k), [`DeepSeek-Coder`](https://github.com/deepseek-ai/DeepSeek-Coder) (23.6k), and [`DeepSeek-OCR`](https://github.com/deepseek-ai/DeepSeek-OCR) (23.3k). The infra stack is its own cluster: [`Janus`](https://github.com/deepseek-ai/Janus) (17.7k), [`FlashMLA`](https://github.com/deepseek-ai/FlashMLA) (12.7k), [`3FS`](https://github.com/deepseek-ai/3FS) (10.0k), [`DeepEP`](https://github.com/deepseek-ai/DeepEP) (9.7k), [`open-infra-index`](https://github.com/deepseek-ai/open-infra-index) (8.0k), [`DeepGEMM`](https://github.com/deepseek-ai/DeepGEMM) (7.4k), and [`DeepSeek-LLM`](https://github.com/deepseek-ai/DeepSeek-LLM) (7.0k). Recent tagged releases skew toward infra cadence: multiple [DeepGEMM](https://github.com/deepseek-ai/DeepGEMM/releases/tag/v2.1.1.post3) builds (a string of `nv_dev_*` dev tags plus `v2.1.1.post3`) and [DeepEP `v1.2.1`](https://github.com/deepseek-ai/DeepEP/releases/tag/v1.2.1), alongside the `v1.0.0` tags for [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3/releases/tag/v1.0.0) and [DeepSeek-R1](https://github.com/deepseek-ai/DeepSeek-R1/releases/tag/v1.0.0).\n\n## Research themes\n\nNo first-party writing captured yet.\n\n## Hiring & scaling\n\nNo careers data captured yet.\n\n## Traction highlights\n\nHacker News attention is concentrated on a single launch: [`DeepSeek-OCR`](https://github.com/deepseek-ai/DeepSeek-OCR) drew **1,003 points / 244 comments**, an order of magnitude above the next thread, [`DeepSeek-V3.2-Exp`](https://github.com/deepseek-ai/DeepSeek-V3.2-Exp) at **309 points / 50 comments**. A long tail of smaller/newer repos also surfaced — [`LPLB`](https://github.com/deepseek-ai/LPLB) (43 pts), [`DeepSeek-Math-V2`](https://github.com/deepseek-ai/DeepSeek-Math-V2) (9 pts), [`DeepSeek-OCR-2`](https://github.com/deepseek-ai/DeepSeek-OCR-2) (7 pts), [`Engram`](https://github.com/deepseek-ai/Engram) (3 pts), and [`TileKernels`](https://github.com/deepseek-ai/TileKernels) (2 pts). On adoption, the most-downloaded model is [`DeepSeek-R1-0528`](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528) (6.44M/30d) and the most-starred repo is [`DeepSeek-V3`](https://github.com/deepseek-ai/DeepSeek-V3) (103.7k stars).\n\n## Sources\n\n- https://www.deepseek.com/\n- https://huggingface.co/deepseek-ai/DeepSeek-R1-0528\n- https://huggingface.co/deepseek-ai/DeepSeek-R1\n- https://huggingface.co/deepseek-ai/DeepSeek-V3.2\n- https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp\n- https://huggingface.co/deepseek-ai/DeepSeek-OCR\n- https://huggingface.co/deepseek-ai/DeepSeek-OCR-2\n- https://huggingface.co/deepseek-ai/DeepSeek-V3\n- https://huggingface.co/deepseek-ai/deepseek-vl2-tiny\n- https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B\n- https://github.com/deepseek-ai/DeepSeek-V3\n- https://github.com/deepseek-ai/DeepSeek-R1\n- https://github.com/deepseek-ai/awesome-deepseek-integration\n- https://github.com/deepseek-ai/DeepSeek-Coder\n- https://github.com/deepseek-ai/DeepSeek-OCR\n- https://github.com/deepseek-ai/Janus\n- https://github.com/deepseek-ai/FlashMLA\n- https://github.com/deepseek-ai/3FS\n- https://github.com/deepseek-ai/DeepEP\n- https://github.com/deepseek-ai/DeepGEMM\n- https://github.com/deepseek-ai/open-infra-index\n- https://github.com/deepseek-ai/DeepGEMM/releases/tag/v2.1.1.post3\n- https://github.com/deepseek-ai/DeepEP/releases/tag/v1.2.1\n- https://github.com/deepseek-ai/LPLB\n- https://github.com/deepseek-ai/DeepSeek-Math-V2","generated_at":"2026-06-08T15:59:08.656+00:00","citations":[{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1","path":null,"label":"deepseek-ai/DeepSeek-R1","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-V3","path":null,"label":"deepseek-ai/DeepSeek-V3","type":"external"},{"url":"https://github.com/deepseek-ai/DeepGEMM","path":null,"label":"deepseek-ai/DeepGEMM","type":"external"},{"url":"https://github.com/deepseek-ai/DeepEP","path":null,"label":"deepseek-ai/DeepEP","type":"external"},{"url":"https://github.com/deepseek-ai/FlashMLA","path":null,"label":"deepseek-ai/FlashMLA","type":"external"},{"url":"https://github.com/deepseek-ai/3FS","path":null,"label":"deepseek-ai/3FS","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-0528","path":null,"label":"deepseek-ai/DeepSeek-R1-0528","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-V3.2","path":null,"label":"deepseek-ai/DeepSeek-V3.2","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp","path":null,"label":"deepseek-ai/DeepSeek-V3.2-Exp","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-OCR","path":null,"label":"deepseek-ai/DeepSeek-OCR","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-OCR-2","path":null,"label":"deepseek-ai/DeepSeek-OCR-2","type":"external"},{"url":"https://huggingface.co/deepseek-ai/deepseek-vl2-tiny","path":null,"label":"deepseek-ai/deepseek-vl2-tiny","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B","path":null,"label":"deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B","path":null,"label":"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B","path":null,"label":"deepseek-ai/DeepSeek-R1-Distill-Qwen-14B","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B","path":null,"label":"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B","path":null,"label":"deepseek-ai/DeepSeek-R1-Distill-Llama-8B","type":"external"},{"url":"https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B","path":null,"label":"deepseek-ai/DeepSeek-R1-0528-Qwen3-8B","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-V3","path":null,"label":"deepseek-ai/DeepSeek-V3","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-R1","path":null,"label":"deepseek-ai/DeepSeek-R1","type":"external"},{"url":"https://github.com/deepseek-ai/awesome-deepseek-integration","path":null,"label":"deepseek-ai/awesome-deepseek-integration","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-Coder","path":null,"label":"deepseek-ai/DeepSeek-Coder","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-OCR","path":null,"label":"deepseek-ai/DeepSeek-OCR","type":"external"},{"url":"https://github.com/deepseek-ai/Janus","path":null,"label":"deepseek-ai/Janus","type":"external"},{"url":"https://github.com/deepseek-ai/open-infra-index","path":null,"label":"deepseek-ai/open-infra-index","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-LLM","path":null,"label":"deepseek-ai/DeepSeek-LLM","type":"external"},{"url":"https://github.com/deepseek-ai/DeepGEMM/releases/tag/v2.1.1.post3","path":null,"label":"deepseek-ai/DeepGEMM","type":"external"},{"url":"https://github.com/deepseek-ai/DeepEP/releases/tag/v1.2.1","path":null,"label":"deepseek-ai/DeepEP","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-V3/releases/tag/v1.0.0","path":null,"label":"deepseek-ai/DeepSeek-V3","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-R1/releases/tag/v1.0.0","path":null,"label":"deepseek-ai/DeepSeek-R1","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-V3.2-Exp","path":null,"label":"deepseek-ai/DeepSeek-V3.2-Exp","type":"external"},{"url":"https://github.com/deepseek-ai/LPLB","path":null,"label":"deepseek-ai/LPLB","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-Math-V2","path":null,"label":"deepseek-ai/DeepSeek-Math-V2","type":"external"},{"url":"https://github.com/deepseek-ai/DeepSeek-OCR-2","path":null,"label":"deepseek-ai/DeepSeek-OCR-2","type":"external"},{"url":"https://github.com/deepseek-ai/Engram","path":null,"label":"deepseek-ai/Engram","type":"external"},{"url":"https://github.com/deepseek-ai/TileKernels","path":null,"label":"deepseek-ai/TileKernels","type":"external"},{"url":"https://www.deepseek.com/","path":null,"label":"deepseek.com","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"cohere","url":"https://onlylabs.fyi/analysis/cohere","json_url":"https://onlylabs.fyi/analysis/cohere/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/cohere/evidence.json","dossier_url":"https://onlylabs.fyi/labs/cohere","org":{"slug":"cohere","name":"Cohere","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://cohere.com"},"title":"Cohere analysis","summary":"Cohere is an enterprise-focused LLM provider whose public footprint splits cleanly in two: a commercial Command line of general-purpose models and the open-weight, multilingual research effort shipped under the Cohere Labs org, centered on the Aya family. Recent open releases — command-a-plus-05-2026-bf16, command-a-vision-07-2025, and a new cohere-transcribe-03-2026 speech model — show the lab expanding from text…","markdown":"## Thesis\n\nCohere is an enterprise-focused LLM provider whose public footprint splits cleanly in two: a commercial Command line of general-purpose models and the open-weight, multilingual research effort shipped under the Cohere Labs org, centered on the Aya family. Recent open releases — [`command-a-plus-05-2026-bf16`](https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16), [`command-a-vision-07-2025`](https://huggingface.co/CohereLabs/command-a-vision-07-2025), and a new [`cohere-transcribe-03-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026) speech model — show the lab expanding from text into vision and audio while doubling down on multilingual coverage. Its hiring is heavily go-to-market and security-/clearance-oriented, consistent with a company selling into regulated enterprise and government buyers.\n\n## Shipping\n\nCohere ships open weights under the `CohereLabs` org on Hugging Face. By recent 30-day downloads:\n\n- [`cohere-transcribe-03-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026) — 509,247 downloads, 977 likes (~2.1B params), the lab's top puller and a move into speech/transcription.\n- [`aya-vision-8b`](https://huggingface.co/CohereLabs/aya-vision-8b) — 172,612 downloads, 321 likes (~8.6B params), multilingual vision.\n- [`command-a-vision-07-2025`](https://huggingface.co/CohereLabs/command-a-vision-07-2025) — 118,386 downloads, 89 likes (~112B params), the vision-enabled Command A.\n- [`c4ai-command-r7b-arabic-02-2025`](https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025) — 51,610 downloads, a ~8B model targeting Arabic.\n- [`aya-expanse-8b`](https://huggingface.co/CohereLabs/aya-expanse-8b) — 39,401 downloads, 434 likes; [`aya-expanse-32b`](https://huggingface.co/CohereLabs/aya-expanse-32b) — 12,434 downloads.\n- [`command-a-plus-05-2026-bf16`](https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16) — 21,233 downloads (~219B params), the largest and most recent Command release in the set.\n- Smaller/efficiency entries: [`tiny-aya-global`](https://huggingface.co/CohereLabs/tiny-aya-global) (23,867) and [`tiny-aya-base`](https://huggingface.co/CohereLabs/tiny-aya-base) (16,980), both ~3.3B.\n\nThe most \"liked\" weights are older flagships — [`c4ai-command-r-plus`](https://huggingface.co/CohereLabs/c4ai-command-r-plus) (1,794 likes, ~104B) and [`c4ai-command-r-v01`](https://huggingface.co/CohereLabs/c4ai-command-r-v01) (1,109 likes, ~35B) — which still draw 4,913 and 31,337 downloads respectively, indicating durable community attachment to the original Command R line.\n\nOn the tooling side, the marquee repo is [`cohere-ai/cohere-toolkit`](https://github.com/cohere-ai/cohere-toolkit) at 3,179 stars, backed by official SDKs ([`cohere-python`](https://github.com/cohere-ai/cohere-python), 388 stars; [`cohere-typescript`](https://github.com/cohere-ai/cohere-typescript), 173 stars). Tagged releases in the window are mostly infrastructure and SDK housekeeping — repeated [`cloud-api-adaptor` v0.2.0-cohere.x](https://github.com/cohere-ai/cloud-api-adaptor/releases/tag/v0.2.0-cohere.5) cuts, [`cohere-python` 7.0.3 / 7.0.2](https://github.com/cohere-ai/cohere-python/releases/tag/7.0.3), and an early [`TNG` v0.1.0](https://github.com/cohere-ai/TNG/releases/tag/v0.1.0) — rather than model drops.\n\n## Research themes\n\nNo first-party writing captured yet. (Direction is inferable from shipped artifacts: multilingual coverage via the Aya / Aya Expanse line and Arabic-specific Command R7B, multimodality via Aya Vision and Command A Vision, speech via Cohere Transcribe, and retrieval tooling via [`BinaryVectorDB`](https://github.com/cohere-ai/BinaryVectorDB) and [`DiskVectorIndex`](https://github.com/cohere-ai/DiskVectorIndex) — but no posts or Pages are in the context to cite directly.)\n\n## Hiring & scaling\n\nThe 15 open roles skew strongly toward commercialization and security rather than core model research:\n\n- **Go-to-market / revenue** dominates: Account Executive (Singapore), Senior Manager Inside Sales (Toronto), Head of Sales Enablement (San Francisco), Senior Manager People Operations (Europe), plus two RevOps roles in New York (Analytics Analyst, GTM Systems Architect) and two Solutions Architecture roles (Solutions Architect, Saudi Arabia; Head of Solutions Architecture, San Francisco). The Saudi Arabia and Singapore postings point at active international/enterprise expansion.\n- **Security and compliance** is unusually heavy for a model lab: Chief Information Security Officer (Toronto), Manager Security Engineering (US), and notably an Infrastructure Security Engineer requiring \"Secret + Clearance\" (Toronto) — a clear signal of government / defense / regulated-industry sales motion.\n- **Technical staff** roles are narrower and pointed: Member of Technical Staff, Multilingual (London) and Senior Member of Technical Staff, Web Data (Toronto) align with the multilingual + data-pipeline emphasis seen in the models, and a Senior Search Applications Performance Engineer (Toronto) reflects the retrieval/search product line. A Staff UX Researcher (Toronto) rounds out product investment.\n\nNet: the headcount story is about selling and securing an enterprise platform — anchored in Toronto with SF, London, NY, Singapore, and Saudi Arabia nodes — more than about scaling a research org.\n\n## Traction highlights\n\nNo Hacker News threads are present in the context (`traction` is empty). Traction is therefore read off open-source and Hugging Face signals:\n\n- **Most-starred repo:** [`cohere-ai/cohere-toolkit`](https://github.com/cohere-ai/cohere-toolkit) at 3,179 stars, well ahead of the next repos ([`notebooks`](https://github.com/cohere-ai/notebooks) 506, [`cohere-python`](https://github.com/cohere-ai/cohere-python) 388, [`cohere-terrarium`](https://github.com/cohere-ai/cohere-terrarium) 314).\n- **Most-downloaded model:** [`cohere-transcribe-03-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026), 509,247 downloads / 977 likes — roughly 3x the next model, [`aya-vision-8b`](https://huggingface.co/CohereLabs/aya-vision-8b) (172,612).\n- **Most-liked weights:** [`c4ai-command-r-plus`](https://huggingface.co/CohereLabs/c4ai-command-r-plus) (1,794) and [`c4ai-command-r-v01`](https://huggingface.co/CohereLabs/c4ai-command-r-v01) (1,109), the highest community engagement despite lower current download volume.\n\n## Sources\n\n- Homepage: https://cohere.com\n- [`CohereLabs/cohere-transcribe-03-2026`](https://huggingface.co/CohereLabs/cohere-transcribe-03-2026)\n- [`CohereLabs/aya-vision-8b`](https://huggingface.co/CohereLabs/aya-vision-8b)\n- [`CohereLabs/command-a-vision-07-2025`](https://huggingface.co/CohereLabs/command-a-vision-07-2025)\n- [`CohereLabs/c4ai-command-r7b-arabic-02-2025`](https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025)\n- [`CohereLabs/aya-expanse-8b`](https://huggingface.co/CohereLabs/aya-expanse-8b) · [`aya-expanse-32b`](https://huggingface.co/CohereLabs/aya-expanse-32b)\n- [`CohereLabs/command-a-plus-05-2026-bf16`](https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16)\n- [`CohereLabs/c4ai-command-r-plus`](https://huggingface.co/CohereLabs/c4ai-command-r-plus) · [`c4ai-command-r-v01`](https://huggingface.co/CohereLabs/c4ai-command-r-v01)\n- [`CohereLabs/tiny-aya-global`](https://huggingface.co/CohereLabs/tiny-aya-global) · [`tiny-aya-base`](https://huggingface.co/CohereLabs/tiny-aya-base)\n- [`cohere-ai/cohere-toolkit`](https://github.com/cohere-ai/cohere-toolkit)\n- [`cohere-ai/cohere-python`](https://github.com/cohere-ai/cohere-python) · [release 7.0.3](https://github.com/cohere-ai/cohere-python/releases/tag/7.0.3)\n- [`cohere-ai/cohere-typescript`](https://github.com/cohere-ai/cohere-typescript)\n- [`cohere-ai/BinaryVectorDB`](https://github.com/cohere-ai/BinaryVectorDB) · [`DiskVectorIndex`](https://github.com/cohere-ai/DiskVectorIndex)\n- [`cohere-ai/cloud-api-adaptor` v0.2.0-cohere.5](https://github.com/cohere-ai/cloud-api-adaptor/releases/tag/v0.2.0-cohere.5)\n- [`cohere-ai/TNG` v0.1.0](https://github.com/cohere-ai/TNG/releases/tag/v0.1.0)","generated_at":"2026-06-08T15:59:08.543+00:00","citations":[{"url":"https://huggingface.co/CohereLabs/command-a-plus-05-2026-bf16","path":null,"label":"CohereLabs/command-a-plus-05-2026-bf16","type":"external"},{"url":"https://huggingface.co/CohereLabs/command-a-vision-07-2025","path":null,"label":"CohereLabs/command-a-vision-07-2025","type":"external"},{"url":"https://huggingface.co/CohereLabs/cohere-transcribe-03-2026","path":null,"label":"CohereLabs/cohere-transcribe-03-2026","type":"external"},{"url":"https://huggingface.co/CohereLabs/aya-vision-8b","path":null,"label":"CohereLabs/aya-vision-8b","type":"external"},{"url":"https://huggingface.co/CohereLabs/c4ai-command-r7b-arabic-02-2025","path":null,"label":"CohereLabs/c4ai-command-r7b-arabic-02-2025","type":"external"},{"url":"https://huggingface.co/CohereLabs/aya-expanse-8b","path":null,"label":"CohereLabs/aya-expanse-8b","type":"external"},{"url":"https://huggingface.co/CohereLabs/aya-expanse-32b","path":null,"label":"CohereLabs/aya-expanse-32b","type":"external"},{"url":"https://huggingface.co/CohereLabs/tiny-aya-global","path":null,"label":"CohereLabs/tiny-aya-global","type":"external"},{"url":"https://huggingface.co/CohereLabs/tiny-aya-base","path":null,"label":"CohereLabs/tiny-aya-base","type":"external"},{"url":"https://huggingface.co/CohereLabs/c4ai-command-r-plus","path":null,"label":"CohereLabs/c4ai-command-r-plus","type":"external"},{"url":"https://huggingface.co/CohereLabs/c4ai-command-r-v01","path":null,"label":"CohereLabs/c4ai-command-r-v01","type":"external"},{"url":"https://github.com/cohere-ai/cohere-toolkit","path":null,"label":"cohere-ai/cohere-toolkit","type":"external"},{"url":"https://github.com/cohere-ai/cohere-python","path":null,"label":"cohere-ai/cohere-python","type":"external"},{"url":"https://github.com/cohere-ai/cohere-typescript","path":null,"label":"cohere-ai/cohere-typescript","type":"external"},{"url":"https://github.com/cohere-ai/cloud-api-adaptor/releases/tag/v0.2.0-cohere.5","path":null,"label":"cohere-ai/cloud-api-adaptor","type":"external"},{"url":"https://github.com/cohere-ai/cohere-python/releases/tag/7.0.3","path":null,"label":"cohere-ai/cohere-python","type":"external"},{"url":"https://github.com/cohere-ai/TNG/releases/tag/v0.1.0","path":null,"label":"cohere-ai/TNG","type":"external"},{"url":"https://github.com/cohere-ai/BinaryVectorDB","path":null,"label":"cohere-ai/BinaryVectorDB","type":"external"},{"url":"https://github.com/cohere-ai/DiskVectorIndex","path":null,"label":"cohere-ai/DiskVectorIndex","type":"external"},{"url":"https://github.com/cohere-ai/notebooks","path":null,"label":"cohere-ai/notebooks","type":"external"},{"url":"https://github.com/cohere-ai/cohere-terrarium","path":null,"label":"cohere-ai/cohere-terrarium","type":"external"},{"url":"https://cohere.com","path":null,"label":"cohere.com","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"bytedance","url":"https://onlylabs.fyi/analysis/bytedance","json_url":"https://onlylabs.fyi/analysis/bytedance/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/bytedance/evidence.json","dossier_url":"https://onlylabs.fyi/labs/bytedance","org":{"slug":"bytedance","name":"ByteDance (Doubao/Seed)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://seed.bytedance.com/en/"},"title":"ByteDance (Doubao/Seed) analysis","summary":"ByteDance's Seed lab (Doubao/Seed) is shipping aggressively in the open, spanning a foundation-model line (Seed-OSS), specialized reasoning/coding/proving models, multimodal and 3D-vision systems, and the distributed-training and GPU-kernel infrastructure underneath them. The footprint reads as a full-stack research org: it open-sources both the models and the training/serving substrate (VeOmni, Triton-distributed)…","markdown":"## Thesis\n\nByteDance's Seed lab (Doubao/Seed) is shipping aggressively in the open, spanning a foundation-model line (Seed-OSS), specialized reasoning/coding/proving models, multimodal and 3D-vision systems, and the distributed-training and GPU-kernel infrastructure underneath them. The footprint reads as a full-stack research org: it open-sources both the models and the training/serving substrate (VeOmni, Triton-distributed) rather than weights alone. Breadth across modalities — language, vision-language, video, depth, code, theorem proving — is the defining signal.\n\n## Shipping\n\nThe flagship open release is **[Seed-OSS-36B-Instruct](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct)** (~36B params), the lab's most-downloaded model at **39,280** 30-day downloads and **499** likes, paired with **[Seed-OSS-36B-Base](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base)** (1,852 downloads) and an ablation variant **[Seed-OSS-36B-Base-woSyn](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base-woSyn)** (116 downloads, 53 likes). The supporting repo **[seed-oss](https://github.com/ByteDance-Seed/seed-oss)** has 885 stars.\n\nBeyond the base line, the lab ships narrowly specialized models:\n- **[Stable-DiffCoder-8B-Instruct](https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct)** (766 downloads, 137 likes) and its **[Base](https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Base)** (389 downloads) — diffusion-style code generation.\n- **[BFS-Prover-V2-7B](https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-7B)** (606 downloads) and **[BFS-Prover-V2-32B](https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-32B)** (224 downloads) — automated theorem proving.\n- **[cudaLLM-8B](https://huggingface.co/ByteDance-Seed/cudaLLM-8B)** (178 downloads, 29 likes) — CUDA-kernel generation.\n- A family of **AHN** (Artificial Hippocampus Network) adapters for Qwen-2.5-Instruct at 3B/7B/14B in Mamba2, GDN, and DN variants (each ~50–66 downloads), e.g. **[AHN-Mamba2-for-Qwen-2.5-Instruct-3B](https://huggingface.co/ByteDance-Seed/AHN-Mamba2-for-Qwen-2.5-Instruct-3B)**.\n- **[M3-Agent-Control](https://huggingface.co/ByteDance-Seed/M3-Agent-Control)** (~33B, 50 downloads / 50 likes), the control model behind the m3-agent line.\n\nOn GitHub the most-starred work skews multimodal and vision: **[Bagel](https://github.com/ByteDance-Seed/Bagel)** (5,994 stars), **[Depth-Anything-3](https://github.com/ByteDance-Seed/Depth-Anything-3)** (5,489), **[Seed1.5-VL](https://github.com/ByteDance-Seed/Seed1.5-VL)** (1,580), **[m3-agent](https://github.com/ByteDance-Seed/m3-agent)** (1,369), **[SeedVR](https://github.com/ByteDance-Seed/SeedVR)** (1,213), **[VideoWorld](https://github.com/ByteDance-Seed/VideoWorld)** (790), **[Seed-Coder](https://github.com/ByteDance-Seed/Seed-Coder)** (755), **[Seed-Thinking-v1.5](https://github.com/ByteDance-Seed/Seed-Thinking-v1.5)** (812), and **[TraceAnything](https://github.com/ByteDance-Seed/TraceAnything)** (540).\n\nInfrastructure ships on its own cadence: the training framework **[VeOmni](https://github.com/ByteDance-Seed/VeOmni)** (1,991 stars) has the most active release stream — through **[v0.1.11](https://github.com/ByteDance-Seed/VeOmni/releases/tag/v0.1.11)** (plus v0.1.10 and several v0.1.9 alphas) — and **[Triton-distributed](https://github.com/ByteDance-Seed/Triton-distributed)** (1,455 stars) has cut tagged releases from `experimental` through **[v0.0.2-rc](https://github.com/ByteDance-Seed/Triton-distributed/releases/tag/v0.0.2-rc)**. Smaller tooling releases include **[JoltQC v0.1](https://github.com/ByteDance-Seed/JoltQC/releases/tag/v0.1)** and a **[bamboo_mixer manuscript](https://github.com/ByteDance-Seed/bamboo_mixer/releases/tag/manuscript)** tag.\n\n## Research themes\n\nNo first-party writing captured yet. (Directions are nonetheless legible from the shipped artifacts: multimodal understanding/generation — Bagel, Seed1.5-VL; 3D and video perception — Depth-Anything-3, SeedVR, VideoWorld, TraceAnything; agents with memory — m3-agent / M3-Agent-Control and the AHN \"Artificial Hippocampus\" adapters; code and CUDA generation — Seed-Coder, Stable-DiffCoder, cudaLLM; theorem proving — BFS-Prover; and reasoning — Seed-Thinking-v1.5.)\n\n## Hiring & scaling\n\nNo careers data captured yet.\n\n## Traction highlights\n\n- **Most-starred repos:** [Bagel](https://github.com/ByteDance-Seed/Bagel) (5,994 stars) and [Depth-Anything-3](https://github.com/ByteDance-Seed/Depth-Anything-3) (5,489) lead by a wide margin, followed by the [VeOmni](https://github.com/ByteDance-Seed/VeOmni) training framework (1,991) and [Seed1.5-VL](https://github.com/ByteDance-Seed/Seed1.5-VL) (1,580).\n- **Most-downloaded models:** [Seed-OSS-36B-Instruct](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct) dominates at 39,280 30-day downloads (499 likes) — roughly 21× the next model, [Seed-OSS-36B-Base](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base) (1,852).\n- **Likes-to-downloads outliers:** [Stable-DiffCoder-8B-Instruct](https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct) (137 likes on 766 downloads) and [M3-Agent-Control](https://huggingface.co/ByteDance-Seed/M3-Agent-Control) (50 likes on 50 downloads) draw outsized interest relative to volume.\n- No Hacker News threads captured yet.\n\n## Sources\n\n- Lab homepage: https://seed.bytedance.com/en/\n- https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct\n- https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base\n- https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct\n- https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-7B\n- https://huggingface.co/ByteDance-Seed/cudaLLM-8B\n- https://huggingface.co/ByteDance-Seed/M3-Agent-Control\n- https://github.com/ByteDance-Seed/Bagel\n- https://github.com/ByteDance-Seed/Depth-Anything-3\n- https://github.com/ByteDance-Seed/VeOmni (releases: https://github.com/ByteDance-Seed/VeOmni/releases/tag/v0.1.11)\n- https://github.com/ByteDance-Seed/Seed1.5-VL\n- https://github.com/ByteDance-Seed/Triton-distributed (releases: https://github.com/ByteDance-Seed/Triton-distributed/releases/tag/v0.0.2-rc)\n- https://github.com/ByteDance-Seed/m3-agent\n- https://github.com/ByteDance-Seed/seed-oss\n- https://github.com/ByteDance-Seed/Seed-Coder\n- https://github.com/ByteDance-Seed/Seed-Thinking-v1.5","generated_at":"2026-06-08T15:59:08.419+00:00","citations":[{"url":"https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct","path":null,"label":"ByteDance-Seed/Seed-OSS-36B-Instruct","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base","path":null,"label":"ByteDance-Seed/Seed-OSS-36B-Base","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base-woSyn","path":null,"label":"ByteDance-Seed/Seed-OSS-36B-Base-woSyn","type":"external"},{"url":"https://github.com/ByteDance-Seed/seed-oss","path":null,"label":"ByteDance-Seed/seed-oss","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct","path":null,"label":"ByteDance-Seed/Stable-DiffCoder-8B-Instruct","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Base","path":null,"label":"ByteDance-Seed/Stable-DiffCoder-8B-Base","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-7B","path":null,"label":"ByteDance-Seed/BFS-Prover-V2-7B","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-32B","path":null,"label":"ByteDance-Seed/BFS-Prover-V2-32B","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/cudaLLM-8B","path":null,"label":"ByteDance-Seed/cudaLLM-8B","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/AHN-Mamba2-for-Qwen-2.5-Instruct-3B","path":null,"label":"ByteDance-Seed/AHN-Mamba2-for-Qwen-2.5-Instruct-3B","type":"external"},{"url":"https://huggingface.co/ByteDance-Seed/M3-Agent-Control","path":null,"label":"ByteDance-Seed/M3-Agent-Control","type":"external"},{"url":"https://github.com/ByteDance-Seed/Bagel","path":null,"label":"ByteDance-Seed/Bagel","type":"external"},{"url":"https://github.com/ByteDance-Seed/Depth-Anything-3","path":null,"label":"ByteDance-Seed/Depth-Anything-3","type":"external"},{"url":"https://github.com/ByteDance-Seed/Seed1.5-VL","path":null,"label":"ByteDance-Seed/Seed1.5-VL","type":"external"},{"url":"https://github.com/ByteDance-Seed/m3-agent","path":null,"label":"ByteDance-Seed/m3-agent","type":"external"},{"url":"https://github.com/ByteDance-Seed/SeedVR","path":null,"label":"ByteDance-Seed/SeedVR","type":"external"},{"url":"https://github.com/ByteDance-Seed/VideoWorld","path":null,"label":"ByteDance-Seed/VideoWorld","type":"external"},{"url":"https://github.com/ByteDance-Seed/Seed-Coder","path":null,"label":"ByteDance-Seed/Seed-Coder","type":"external"},{"url":"https://github.com/ByteDance-Seed/Seed-Thinking-v1.5","path":null,"label":"ByteDance-Seed/Seed-Thinking-v1.5","type":"external"},{"url":"https://github.com/ByteDance-Seed/TraceAnything","path":null,"label":"ByteDance-Seed/TraceAnything","type":"external"},{"url":"https://github.com/ByteDance-Seed/VeOmni","path":null,"label":"ByteDance-Seed/VeOmni","type":"external"},{"url":"https://github.com/ByteDance-Seed/VeOmni/releases/tag/v0.1.11","path":null,"label":"ByteDance-Seed/VeOmni","type":"external"},{"url":"https://github.com/ByteDance-Seed/Triton-distributed","path":null,"label":"ByteDance-Seed/Triton-distributed","type":"external"},{"url":"https://github.com/ByteDance-Seed/Triton-distributed/releases/tag/v0.0.2-rc","path":null,"label":"ByteDance-Seed/Triton-distributed","type":"external"},{"url":"https://github.com/ByteDance-Seed/JoltQC/releases/tag/v0.1","path":null,"label":"ByteDance-Seed/JoltQC","type":"external"},{"url":"https://github.com/ByteDance-Seed/bamboo_mixer/releases/tag/manuscript","path":null,"label":"ByteDance-Seed/bamboo_mixer","type":"external"},{"url":"https://seed.bytedance.com/en/","path":null,"label":"seed.bytedance.com/en","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}},{"org_slug":"baidu","url":"https://onlylabs.fyi/analysis/baidu","json_url":"https://onlylabs.fyi/analysis/baidu/analysis.json","evidence_json_url":"https://onlylabs.fyi/analysis/baidu/evidence.json","dossier_url":"https://onlylabs.fyi/labs/baidu","org":{"slug":"baidu","name":"Baidu (ERNIE)","category":"frontier-lab","category_label":"Frontier lab","homepage_url":"https://ernie.baidu.com"},"title":"Baidu (ERNIE) analysis","summary":"Baidu's public footprint reads as two tightly-coupled engines: the open-weight ERNIE 4.5 model family (dense 0.3B through a 424B-A47B MoE, with vision-language and \"Thinking\" reasoning variants) shipped on Hugging Face, and the PaddlePaddle open-source stack that serves as its distribution and tooling moat — led by PaddleOCR, by far its largest community asset. Document AI / OCR is the clear commercial center of…","markdown":"## Thesis\n\nBaidu's public footprint reads as two tightly-coupled engines: the open-weight **ERNIE 4.5** model family (dense 0.3B through a 424B-A47B MoE, with vision-language and \"Thinking\" reasoning variants) shipped on Hugging Face, and the **PaddlePaddle** open-source stack that serves as its distribution and tooling moat — led by PaddleOCR, by far its largest community asset. Document AI / OCR is the clear commercial center of gravity: its single most-downloaded model is an OCR model, and its most-active release cadence is on PaddleOCR/PaddleX.\n\n## Shipping\n\n**Models (Hugging Face, 30-day downloads):**\n- [`baidu/Qianfan-OCR`](https://huggingface.co/baidu/Qianfan-OCR) — **204,259** downloads, 1,177 likes (~4.7B params), the most-downloaded model in the set; OCR is the demand center.\n- [`baidu/ERNIE-4.5-VL-28B-A3B-PT`](https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT) — **202,248** downloads; the most-pulled vision-language MoE (~29B params, 3B active).\n- [`baidu/ERNIE-4.5-21B-A3B-PT`](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT) — **66,235** downloads; the workhorse mid-size MoE.\n- [`baidu/ERNIE-Image`](https://huggingface.co/baidu/ERNIE-Image) — **35,776** downloads, 641 likes; image generation.\n- [`baidu/ERNIE-4.5-0.3B-PT`](https://huggingface.co/baidu/ERNIE-4.5-0.3B-PT) — **29,297** downloads; the small dense edge model.\n- Reasoning push: [`baidu/ERNIE-4.5-21B-A3B-Thinking`](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking) (12,856 downloads, **786** likes) and [`baidu/ERNIE-4.5-VL-28B-A3B-Thinking`](https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-Thinking) (539 likes) — high like-to-download ratios suggest strong interest relative to actual pull.\n- Frontier scale, low adoption: [`baidu/ERNIE-4.5-300B-A47B-PT`](https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT) (~300B params, 1,375 downloads) and [`baidu/ERNIE-4.5-VL-424B-A47B-Base-PT`](https://huggingface.co/baidu/ERNIE-4.5-VL-424B-A47B-Base-PT) (~424B params, 142 downloads) — the big MoEs are published but lightly pulled, consistent with their cost.\n\nThe lineup spans a full ladder — 0.3B dense → 21B/28B A3B MoE → 300B/424B A47B MoE — in PT, Base-PT, and Thinking flavors, signaling a deliberate one-family-many-sizes strategy. It also includes a Qianfan VL line ([`baidu/Qianfan-VL-8B`](https://huggingface.co/baidu/Qianfan-VL-8B)) and image variants ([`baidu/ERNIE-Image-Turbo`](https://huggingface.co/baidu/ERNIE-Image-Turbo), [`baidu/ERNIE-Image-Aes`](https://huggingface.co/baidu/ERNIE-Image-Aes)).\n\n**Repos (GitHub stars):**\n- [`PaddlePaddle/PaddleOCR`](https://github.com/PaddlePaddle/PaddleOCR) — **81,367** stars, the dominant asset (PaddleOCR shipped v3.4.1 → v3.6.0 in the recent window).\n- [`PaddlePaddle/Paddle`](https://github.com/PaddlePaddle/Paddle) — 23,938 stars, the core framework.\n- [`PaddlePaddle/PaddleDetection`](https://github.com/PaddlePaddle/PaddleDetection) (14,238), [`PaddlePaddle/PaddleFormers`](https://github.com/PaddlePaddle/PaddleFormers) (12,982), [`PaddlePaddle/PaddleNLP`](https://github.com/PaddlePaddle/PaddleNLP) (12,952), [`PaddlePaddle/PaddleSpeech`](https://github.com/PaddlePaddle/PaddleSpeech) (12,614), [`PaddlePaddle/ERNIE`](https://github.com/PaddlePaddle/ERNIE) (7,722), [`PaddlePaddle/PaddleX`](https://github.com/PaddlePaddle/PaddleX) (6,156).\n\n**Releases:** the recent cadence is heavily PaddleX/PaddleOCR — [PaddleOCR v3.6.0](https://github.com/PaddlePaddle/PaddleOCR/releases/tag/v3.6.0), [PaddleX v3.6.1](https://github.com/PaddlePaddle/PaddleX/releases/tag/v3.6.1) and v3.5.0–v3.6.0, plus [FastDeploy v2.5.0](https://github.com/PaddlePaddle/FastDeploy/releases/tag/v2.5.0) (inference deployment) and [PaddleFormers 1.1.1](https://github.com/PaddlePaddle/PaddleFormers/releases/tag/1.1.1) — i.e. shipping is concentrated on the OCR product surface and the deployment/training tooling around it.\n\n## Research themes\n\nNo first-party writing captured yet.\n\n## Hiring & scaling\n\nOpen roles are entirely US-based (Mountain View and Sunnyvale, CA) and split into two clear buckets:\n\n- **Silicon / ML systems** in Sunnyvale: *Machine Learning System Hardware Architect*, *Machine Learning System Software Architect*, *CPU Digital Design Engineer*, *CPU/GPU/Processor Hardware Architect*, and *Design Verification Engineer*. This is a hardware-and-systems hiring cluster — chip-level design and ML-systems architecture, pointing at in-house compute/accelerator investment rather than model research.\n- **Commercial / GTM** in Mountain View: *GTM Strategy & Operations (AI Desktop & Mobile App)*, *Senior Business Development & Partnerships Manager*, *Advertising Sales Manager*, multiple *Account Manager* and sales/campaign roles, and *Head of Global Business Unit, Baidu USA*. The \"AI Desktop & Mobile App\" GTM role plus the ad-sales and BD weighting signal a US consumer-app monetization and partnerships push.\n\nThere is also an *R&D Software Engineer Lead & CISO* (Mountain View). Notably, the captured roles contain no model-research/applied-science postings — the visible US hiring is hardware-systems + go-to-market, not core model R&D.\n\n## Traction highlights\n\n- **Most-starred repo:** [`PaddlePaddle/PaddleOCR`](https://github.com/PaddlePaddle/PaddleOCR) at **81,367** stars — the standout, several times larger than the core [Paddle](https://github.com/PaddlePaddle/Paddle) framework (23,938).\n- **Most-downloaded model:** [`baidu/Qianfan-OCR`](https://huggingface.co/baidu/Qianfan-OCR) at **204,259** 30-day downloads, narrowly ahead of [`ERNIE-4.5-VL-28B-A3B-PT`](https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT) (202,248).\n- **Hacker News:** thin traction — the top thread is [`PaddlePaddle/PaddleOCR`](https://github.com/PaddlePaddle/PaddleOCR) at **19 points / 3 comments**, with [`PaddlePaddle/PaddleSeg`](https://github.com/PaddlePaddle/PaddleSeg) at just 1 point. Baidu's reach shows up in download/star counts, not in HN discussion.\n- **Most-liked models** skew toward the reasoning variants — [`ERNIE-4.5-21B-A3B-Thinking`](https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking) (786 likes) and [`Qianfan-OCR`](https://huggingface.co/baidu/Qianfan-OCR) (1,177 likes) lead community attention.\n\n## Sources\n\n- ERNIE homepage: https://ernie.baidu.com\n- https://huggingface.co/baidu/Qianfan-OCR\n- https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT\n- https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT\n- https://huggingface.co/baidu/ERNIE-Image\n- https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking\n- https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT\n- https://huggingface.co/baidu/ERNIE-4.5-VL-424B-A47B-Base-PT\n- https://github.com/PaddlePaddle/PaddleOCR\n- https://github.com/PaddlePaddle/Paddle\n- https://github.com/PaddlePaddle/PaddleX\n- https://github.com/PaddlePaddle/PaddleOCR/releases/tag/v3.6.0\n- https://github.com/PaddlePaddle/PaddleX/releases/tag/v3.6.1\n- https://github.com/PaddlePaddle/FastDeploy/releases/tag/v2.5.0\n- https://github.com/PaddlePaddle/PaddleSeg","generated_at":"2026-06-08T15:59:08.147+00:00","citations":[{"url":"https://huggingface.co/baidu/Qianfan-OCR","path":null,"label":"baidu/Qianfan-OCR","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-PT","path":null,"label":"baidu/ERNIE-4.5-VL-28B-A3B-PT","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT","path":null,"label":"baidu/ERNIE-4.5-21B-A3B-PT","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-Image","path":null,"label":"baidu/ERNIE-Image","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-0.3B-PT","path":null,"label":"baidu/ERNIE-4.5-0.3B-PT","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-Thinking","path":null,"label":"baidu/ERNIE-4.5-21B-A3B-Thinking","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-VL-28B-A3B-Thinking","path":null,"label":"baidu/ERNIE-4.5-VL-28B-A3B-Thinking","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT","path":null,"label":"baidu/ERNIE-4.5-300B-A47B-PT","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-4.5-VL-424B-A47B-Base-PT","path":null,"label":"baidu/ERNIE-4.5-VL-424B-A47B-Base-PT","type":"external"},{"url":"https://huggingface.co/baidu/Qianfan-VL-8B","path":null,"label":"baidu/Qianfan-VL-8B","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-Image-Turbo","path":null,"label":"baidu/ERNIE-Image-Turbo","type":"external"},{"url":"https://huggingface.co/baidu/ERNIE-Image-Aes","path":null,"label":"baidu/ERNIE-Image-Aes","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleOCR","path":null,"label":"PaddlePaddle/PaddleOCR","type":"external"},{"url":"https://github.com/PaddlePaddle/Paddle","path":null,"label":"PaddlePaddle/Paddle","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleDetection","path":null,"label":"PaddlePaddle/PaddleDetection","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleFormers","path":null,"label":"PaddlePaddle/PaddleFormers","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleNLP","path":null,"label":"PaddlePaddle/PaddleNLP","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleSpeech","path":null,"label":"PaddlePaddle/PaddleSpeech","type":"external"},{"url":"https://github.com/PaddlePaddle/ERNIE","path":null,"label":"PaddlePaddle/ERNIE","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleX","path":null,"label":"PaddlePaddle/PaddleX","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleOCR/releases/tag/v3.6.0","path":null,"label":"PaddlePaddle/PaddleOCR","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleX/releases/tag/v3.6.1","path":null,"label":"PaddlePaddle/PaddleX","type":"external"},{"url":"https://github.com/PaddlePaddle/FastDeploy/releases/tag/v2.5.0","path":null,"label":"PaddlePaddle/FastDeploy","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleFormers/releases/tag/1.1.1","path":null,"label":"PaddlePaddle/PaddleFormers","type":"external"},{"url":"https://github.com/PaddlePaddle/PaddleSeg","path":null,"label":"PaddlePaddle/PaddleSeg","type":"external"},{"url":"https://ernie.baidu.com","path":null,"label":"ernie.baidu.com","type":"external"}],"provenance":{"provider":null,"model":null,"workflow":"synthesize-analyses","agent":null},"evidence":{"total":null,"pages":null,"events":null,"web":null,"signal_desks":null,"data_radar_lanes":null,"data_radar_matches":null}}]}