{"schema_version":"onlylabs.public_analyses.v1","title":"onlylabs all orgs analysis export","description":"Structured public onlylabs agent analyses: generated markdown reports, cited evidence, provenance, and stable report URLs.","url":"https://onlylabs.fyi/analysis","json_url":"https://onlylabs.fyi/analysis.json","generated_at":"2026-06-13T13:14:47.078Z","scope":{"category":null,"label":"All orgs"},"count":18,"analyses":[{"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 is transitioning from a safety-first research lab into a multi-platform enterprise AI company, executing a high-velocity commercial expansion while preserving its interpretability and alignment research identity. The June 2026 launch of Claude Fable 5 and Claude Mythos 5—a tiered-access model family where the most capable weights are restricted to vetted partners—defines its current posture: ship frontier…","markdown":"```json\n{\n  \"content\": \"## Thesis\\n\\nAnthropic is transitioning from a safety-first research lab into a multi-platform enterprise AI company, executing a high-velocity commercial expansion while preserving its interpretability and alignment research identity. The June 2026 launch of Claude Fable 5 and Claude Mythos 5—a tiered-access model family where the most capable weights are restricted to vetted partners—defines its current posture: ship frontier capability broadly, but gate the highest-risk capabilities behind governance mechanisms [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails)[W6](https://www.youtube.com/watch?v=Y9Wz2PV404E). This is matched by an aggressive GTM buildout spanning sales, customer success, partnerships, and field marketing across North America, EMEA, and APAC [E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008)[E25](https://job-boards.greenhouse.io/anthropic/jobs/5109135008)[E26](https://job-boards.greenhouse.io/anthropic/jobs/5255752008)[E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008)[E34](https://job-boards.greenhouse.io/anthropic/jobs/4989228008)[E35](https://job-boards.greenhouse.io/anthropic/jobs/5076109008). Simultaneously, Anthropic is deepening its agent platform: Claude Code, the Claude Agent SDK (TypeScript and Python), and Claude Code Action are shipping at a rapid cadence, positioning Claude as the runtime for autonomous software engineering [P21](https://github.com/anthropics/claude-code/releases/tag/v2.1.177)[P22](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[P23](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176)[P25](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100)[P26](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177)[P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101). The self-reported statistic that Claude Code generates 4.5% of all public GitHub commits—2.6 million weekly—underscores the scale of this agent footprint [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code). Interpretability research remains a durable differentiator, with ongoing circuits updates, dictionary learning scaling, and sleeper-agent detection work signaling sustained investment in mechanistic understanding [P4](https://www.anthropic.com/research/probes-catch-sleeper-agents)[P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P7](https://www.anthropic.com/research/engineering-challenges-interpretability)[P12](https://www.anthropic.com/research/circuits-updates-august-2024). The thin fork evidence suggests Anthropic primarily builds in-house rather than adapting upstream OSS, with only one detected fork in this pack [E60](https://github.com/anthropics/leptos-chartistry).\\n\\n## Signal desks\\n\\n### Hiring\\n\\n- **Enterprise GTM at scale**: The heaviest hiring concentration is in commercial roles—Engineering Manager, Enterprise [E7](https://job-boards.greenhouse.io/anthropic/jobs/5255912008)[P20](https://job-boards.greenhouse.io/anthropic/jobs/5255912008); Product Manager, Enterprise [E56](https://job-boards.greenhouse.io/anthropic/jobs/5253339008); Manager, Startup Partnerships [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008)[P18](https://job-boards.greenhouse.io/anthropic/jobs/5229558008); Strategic Account Executives in SF and Ontario [E25](https://job-boards.greenhouse.io/anthropic/jobs/5109135008)[E26](https://job-boards.greenhouse.io/anthropic/jobs/5255752008); Account Executive for ASEAN [E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008); Manager, Customer Success in London [E34](https://job-boards.greenhouse.io/anthropic/jobs/4989228008); Partner Enablement Lead for System Integrators [E52](https://job-boards.greenhouse.io/anthropic/jobs/5188391008); Product Marketing Lead for Claude Platform (Cloud) [E27](https://job-boards.greenhouse.io/anthropic/jobs/5198991008); Field Marketing Lead for EMEA [E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008); and Field Marketing Manager [E42](https://job-boards.greenhouse.io/anthropic/jobs/5169167008). This signals a fully-staffed enterprise sales motion targeting both mature and emerging markets.\\n- **Inference infrastructure scaling**: Staff+ Software Engineer for Inference Runtime [E20](https://job-boards.greenhouse.io/anthropic/jobs/5257650008)[P17](https://job-boards.greenhouse.io/anthropic/jobs/5257650008), Staff Software Engineer and Sr. Software Engineer for Inference in London [E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008)[E30](https://job-boards.greenhouse.io/anthropic/jobs/5152348008), and Technical Program Manager for API Platform [E24](https://job-boards.greenhouse.io/anthropic/jobs/5256303008)[P13](https://job-boards.greenhouse.io/anthropic/jobs/5256303008) point to serious investment in the inference serving stack across heterogeneous accelerators (GPUs, TPUs, Trainium) [P17](https://job-boards.greenhouse.io/anthropic/jobs/5257650008).\\n- **Data platform investment**: Product Management for Human Data Platform [E21](https://job-boards.greenhouse.io/anthropic/jobs/5195866008)[P15](https://job-boards.greenhouse.io/anthropic/jobs/5195866008) focuses on labeling interfaces, data quality pipelines, and vendor tooling—indicating that frontier model training still depends heavily on high-quality human data at scale.\\n- **GTM analytics and reporting**: Field Reporting Insights Manager [E51](https://job-boards.greenhouse.io/anthropic/jobs/5253257008)[P1](https://job-boards.greenhouse.io/anthropic/jobs/5253257008) builds certified reporting across Salesforce, Looker, and BigQuery—evidence of a maturing revenue operations function measuring pipeline, forecast, bookings, and rep productivity.\\n- **Safety and policy**: Product Manager for Safeguards Rare Harms [E50](https://job-boards.greenhouse.io/anthropic/jobs/5139628008), Policy Communications Manager [E57](https://job-boards.greenhouse.io/anthropic/jobs/5254582008), Engineering Manager for GRC Platform [E54](https://job-boards.greenhouse.io/anthropic/jobs/4980335008), and Senior/Staff Security Engineer for Threat Intelligence in Zürich [E59](https://job-boards.greenhouse.io/anthropic/jobs/5252342008) show continued investment in safety product and security posture.\\n- **Global hub expansion**: London (Inference Engineers, Field Marketing Lead, Customer Success Manager, Developer Productivity) [E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008)[E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008)[E30](https://job-boards.greenhouse.io/anthropic/jobs/5152348008)[E34](https://job-boards.greenhouse.io/anthropic/jobs/4989228008)[E58](https://job-boards.greenhouse.io/anthropic/jobs/5254803008); Tokyo (Applied AI Architect) [E35](https://job-boards.greenhouse.io/anthropic/jobs/5076109008); Sydney (ASEAN Account Executive) [E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008); Ontario (Strategic Account Executive) [E26](https://job-boards.greenhouse.io/anthropic/jobs/5255752008); Zürich (Threat Intelligence) [E59](https://job-boards.greenhouse.io/anthropic/jobs/5252342008).\\n- **Platform and developer tooling**: IT Systems Engineer for Client Platform Engineering [E22](https://job-boards.greenhouse.io/anthropic/jobs/5255853008)[P14](https://job-boards.greenhouse.io/anthropic/jobs/5255853008) treats the device fleet as infrastructure-as-code; Staff+ Software Engineer for Developer Productivity in London [E58](https://job-boards.greenhouse.io/anthropic/jobs/5254803008); IT Systems Engineer for Enterprise SaaS [E31](https://job-boards.greenhouse.io/anthropic/jobs/5161882008).\\n- **Corporate development**: People Programs M&A Lead [E13](https://job-boards.greenhouse.io/anthropic/jobs/5239794008) suggests Anthropic is building internal capability for acquisitions and integration.\\n- **Procurement at infrastructure scale**: Sr. Manager for Procurement Lease Administration [E16](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[E12](https://job-boards.greenhouse.io/anthropic/jobs/5253835008) manages data center, co-location, GPU, and TPU equipment finance leases—confirming large-scale compute procurement.\\n\\n### Forks\\n\\n- **Single detected fork, low signal**: The only fork in this evidence pack is `anthropics/leptos-chartistry`, forked from `feral-dot-io/leptos-chartistry` [E60](https://github.com/anthropics/leptos-chartistry). Leptos is a Rust web framework and Chartistry is a charting library. This is a minor UI dependency fork with no clear connection to model training, evals, agents, or infrastructure. No other fork activity was cited in this pack.\\n\\n### Releases\\n\\n- **Claude Code rapid iteration**: `anthropics/claude-code` shipped versions v2.1.174 through v2.1.177 in a 24-hour window [E10](https://github.com/anthropics/claude-code/releases/tag/v2.1.177)[E16](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[E41](https://github.com/anthropics/claude-code/releases/tag/v2.1.175)[E46](https://github.com/anthropics/claude-code/releases/tag/v2.1.174), with v2.1.176 including session title localization, Bedrock credential caching improvements, model allowlist enforcement hardening, Linux sandbox symlink fixes, tmux clipboard fixes, Remote Control disconnect fixes, and background agent search improvements [P22](https://github.com/anthropics/claude-code/releases/tag/v2.1.176).\\n- **Agent SDK parity releases**: `claude-agent-sdk-typescript` released v0.3.174 through v0.3.177 [E11](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177)[E17](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176)[E40](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.175)[E45](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.174), and `claude-agent-sdk-python` released v0.2.98 through v0.2.101 [E8](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101)[E14](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100)[E36](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.99)[E43](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.98), both tracking Claude Code CLI parity. Python v0.2.101 added typed `TaskUpdatedMessage` for reliable background task lifecycle tracking [P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101).\\n- **CI/CD agent integration**: `claude-code-action` released v1.0.145 through v1.0.148 [E9](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.148)[E15](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147)[E24](https://job-boards.greenhouse.io/anthropic/jobs/5256303008)[E44](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.145), with v1.0.147 adding a pr-stamp-sweep review workflow [P24](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147).\\n- **Model launch**: Claude Fable 5 and Claude Mythos 5 were released June 9, 2026 [E1](https://www.anthropic.com/news/claude-fable-5-mythos-5)[W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails)[W6](https://www.youtube.com/watch?v=Y9Wz2PV404E). Fable 5 is the widely available tier; Mythos 5 is restricted to vetted partners through Project Glasswing [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails). Pricing: $10/M input tokens, $50/M output tokens [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/).\\n\\n### Talking\\n\\n- **Claude Fable 5/Mythos 5 launch**: The dominant narrative. Two posts—the launch announcement [E1](https://www.anthropic.com/news/claude-fable-5-mythos-5) and the access-control follow-up [E2](https://www.anthropic.com/news/fable-mythos-access)—drew massive attention (2603 and 1306 HN points respectively). The launch frames a tiered-access model: Fable 5 as general availability, Mythos 5 restricted to trusted partners via Project Glasswing [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails)[W6](https://www.youtube.com/watch?v=Y9Wz2PV404E). The Mythos Preview model reportedly discovered thousands of cybersecurity vulnerabilities during testing, which is cited as the reason for restricted release [W6](https://www.youtube.com/watch?v=Y9Wz2PV404E).\\n- **Agent capabilities in science**: Posts on \\\"Making Claude a Chemist\\\" [E3](https://www.anthropic.com/research/making-claude-a-chemist), \\\"Agents in Biology\\\" [E6](https://www.anthropic.com/research/agents-in-biology), and \\\"Coding Agents in the Social Sciences\\\" (referenced across multiple research pages [P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P9](https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions)[P10](https://www.anthropic.com/research/transformer-circuits)) signal a coordinated push to frame Claude as a scientific research agent, not just a coding tool.\\n- **Enterprise ecosystem alliances**: Partnership announcements with TCS [E5](https://www.anthropic.com/news/tcs-anthropic-partnership) and DXC [E53](https://www.anthropic.com/news/dxc-anthropic-alliance), plus the Claude Corps announcement [E4](https://www.anthropic.com/news/claude-corps), show deliberate system-integrator and channel ecosystem construction.\\n- **The Anthropic Institute**: A public research agenda launched May 2026 focused on studying AI's real-world impacts from within a frontier lab—covering economy, science, society, and security [W5](https://www.anthropic.com/research/anthropic-institute-agenda). The \\\"Anthropic Public Record\\\" post [E28](https://www.anthropic.com/news/anthropic-public-record) and \\\"Anthropic Economic Index Survey\\\" (referenced in [P2](https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits)) extend this transparency push.\\n- **Self-improving code narrative**: External coverage reports that Claude writes 80% of Anthropic's code, with the typical engineer merging significantly more code than pre-AI baselines [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code). Claude Code reportedly generates 4.5% of all public GitHub commits [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code).\\n- **AI pause advocacy**: The Anthropic Institute paper includes a call for \\\"a verifiable global mechanism to slow or temporarily pause frontier AI development\\\" [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code)—a notable policy posture from inside a leading lab.\\n- **Interpretability thought leadership**: A sustained stream of research communications spanning mathematical frameworks for transformer circuits [P2](https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits), induction heads [P3](https://www.anthropic.com/research/in-context-learning-and-induction-heads), dictionary learning and monosemanticity [P8](https://www.anthropic.com/research/decomposing-language-models-into-understandable-components), sleeper agent detection [P4](https://www.anthropic.com/research/probes-catch-sleeper-agents)[P11](https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training), discrimination evaluation [P9](https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions), circuits updates [P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P12](https://www.anthropic.com/research/circuits-updates-august-2024), and engineering challenges in scaling interpretability [P7](https://www.anthropic.com/research/engineering-challenges-interpretability)[P10](https://www.anthropic.com/research/transformer-circuits).\\n\\n## Shipping\\n\\nAnthropic's shipping cadence splits into two lanes. First, the **model tier**: Claude Fable 5 and Mythos 5 launched June 9, 2026—a Mythos-class model made generally available with safeguards, with the unrestricted version gated behind Project Glasswing for trusted partners [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails)[W6](https://www.youtube.com/watch?v=Y9Wz2PV404E). Pricing at $10/$50 per million input/output tokens represents a significant reduction from prior Mythos-class pricing [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/). Fable 5 claims state-of-the-art on nearly all tested benchmarks, with particular strength in software engineering (topping SWE-Bench Pro by 11 points) and autonomous long-horizon tasks [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails).\\n\\nSecond, the **agent platform**: Claude Code, the Claude Agent SDK (TypeScript and Python), and Claude Code Action are shipping multiple times daily, with tight version parity across all four artifacts [P21](https://github.com/anthropics/claude-code/releases/tag/v2.1.177)[P22](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[P23](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176)[P25](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100)[P26](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177)[P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101). The SDK enables programmatic agent orchestration with background task lifecycle management [P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101); the GitHub Action integrates agents directly into CI/CD pipelines [P24](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147). The release notes for Claude Code v2.1.176 reveal a maturing enterprise feature set: Bedrock credential caching, model allowlist enforcement, managed settings, and enterprise-oriented fixes [P22](https://github.com/anthropics/claude-code/releases/tag/v2.1.176).\\n\\nNo model weights, datasets, or training infrastructure were released in this evidence pack.\\n\\n## Research themes\\n\\n1. **Mechanistic interpretability at scale**: A multi-year research program spanning mathematical frameworks for transformer circuits [P2](https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits), induction heads [P3](https://www.anthropic.com/research/in-context-learning-and-induction-heads), dictionary learning and monosemanticity [P8](https://www.anthropic.com/research/decomposing-language-models-into-understandable-components), and engineering infrastructure to scale these techniques to production models like Claude 3 Sonnet [P7](https://www.anthropic.com/research/engineering-challenges-interpretability). The Circuits Updates series (April, July, August 2024) [P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P12](https://www.anthropic.com/research/circuits-updates-august-2024) and the qualitative methodology reflections [P10](https://www.anthropic.com/research/transformer-circuits) suggest this is an active, evolving research agenda with regular internal iteration.\\n\\n2. **Deception and alignment**: Sleeper agent research demonstrating that deceptive LLMs can persist through standard safety training [P11](https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training), paired with follow-up work showing simple linear probes can detect defection with >99% AUROC [P4](https://www.anthropic.com/research/probes-catch-sleeper-agents). This directly informs the Mythos/Fable tiered-access architecture.\\n\\n3. **Societal impacts and discrimination**: Empirical evaluation of discrimination in Claude 2.0 decisions across 70 diverse scenarios, with mitigation techniques via prompt engineering [P9](https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions).\\n\\n4. **Scientific agents**: Research positioning Claude as an autonomous agent for chemistry [E3](https://www.anthropic.com/research/making-claude-a-chemist), biology [E6](https://www.anthropic.com/research/agents-in-biology), and social sciences (cited across [P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P9](https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions)[P10](https://www.anthropic.com/research/transformer-circuits)), signaling ambition beyond software engineering into scientific research.\\n\\n5. **AI's macro impact**: The Anthropic Institute's research agenda covers AI's effects on the economy, science, society, and security, with a stated goal of publishing findings for external decision-makers [W5](https://www.anthropic.com/research/anthropic-institute-agenda).\\n\\nEvidence is thin on reinforcement learning research beyond one Code RL hiring signal [E48](https://job-boards.greenhouse.io/anthropic/jobs/5254364008). No cited evidence covers multimodal, vision, or audio research in this pack beyond benchmark claims.\\n\\n## Hiring & scaling\\n\\nAnthropic is hiring across every function of a maturing enterprise AI company. The dominant clusters:\\n\\n- **Enterprise GTM** (15+ distinct roles): Engineering Manager for Enterprise [E7](https://job-boards.greenhouse.io/anthropic/jobs/5255912008)[P20](https://job-boards.greenhouse.io/anthropic/jobs/5255912008), Product Manager for Enterprise [E56](https://job-boards.greenhouse.io/anthropic/jobs/5253339008), Manager of Startup Partnerships [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008)[P18](https://job-boards.greenhouse.io/anthropic/jobs/5229558008), Strategic Account Executives across North America [E25](https://job-boards.greenhouse.io/anthropic/jobs/5109135008)[E26](https://job-boards.greenhouse.io/anthropic/jobs/5255752008), Account Executive for ASEAN [E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008), Manager of Customer Success in London [E34](https://job-boards.greenhouse.io/anthropic/jobs/4989228008), Partner Enablement Lead for System Integrators [E52](https://job-boards.greenhouse.io/anthropic/jobs/5188391008), Product Marketing Lead for Claude Platform (Cloud) [E27](https://job-boards.greenhouse.io/anthropic/jobs/5198991008), Field Marketing Lead for EMEA [E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008), Field Marketing Manager [E42](https://job-boards.greenhouse.io/anthropic/jobs/5169167008), Strategy & Operations Lead for Marketing [E37](https://job-boards.greenhouse.io/anthropic/jobs/5248946008), Product Manager for GTM Experiences [E49](https://job-boards.greenhouse.io/anthropic/jobs/5254623008), Product Support Specialist [E47](https://job-boards.greenhouse.io/anthropic/jobs/4979585008), Web Producer for CMS Publishing [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008)[E18](https://job-boards.greenhouse.io/anthropic/jobs/5257669008), and Field Reporting Insights Manager [E51](https://job-boards.greenhouse.io/anthropic/jobs/5253257008)[P1](https://job-boards.greenhouse.io/anthropic/jobs/5253257008).\\n\\n- **Inference and API infrastructure** (5+ roles): Staff+ Software Engineer for Inference Runtime [E20](https://job-boards.greenhouse.io/anthropic/jobs/5257650008)[P17](https://job-boards.greenhouse.io/anthropic/jobs/5257650008), Staff Software Engineer for Inference in London [E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008), Sr. Software Engineer for Inference in London [E30](https://job-boards.greenhouse.io/anthropic/jobs/5152348008), Technical Program Manager for API Platform [E24](https://job-boards.greenhouse.io/anthropic/jobs/5256303008)[P13](https://job-boards.greenhouse.io/anthropic/jobs/5256303008), and IT Systems Engineer for Enterprise SaaS [E31](https://job-boards.greenhouse.io/anthropic/jobs/5161882008).\\n\\n- **Safety and security** (4+ roles): Product Manager for Safeguards Rare Harms [E50](https://job-boards.greenhouse.io/anthropic/jobs/5139628008), Policy Communications Manager [E57](https://job-boards.greenhouse.io/anthropic/jobs/5254582008), Engineering Manager for GRC Platform [E54](https://job-boards.greenhouse.io/anthropic/jobs/4980335008), and Senior/Staff Security Engineer for Threat Intelligence in Zürich [E59](https://job-boards.greenhouse.io/anthropic/jobs/5252342008).\\n\\n- **Data and platform** (3+ roles): Product Management for Human Data Platform [E21](https://job-boards.greenhouse.io/anthropic/jobs/5195866008)[P15](https://job-boards.greenhouse.io/anthropic/jobs/5195866008), Staff+ Software Engineer for Developer Productivity in London [E58](https://job-boards.greenhouse.io/anthropic/jobs/5254803008), and IT Systems Engineer for Client Platform Engineering [E22](https://job-boards.greenhouse.io/anthropic/jobs/5255853008)[P14](https://job-boards.greenhouse.io/anthropic/jobs/5255853008).\\n\\n- **Corporate infrastructure**: People Programs M&A Lead [E13](https://job-boards.greenhouse.io/anthropic/jobs/5239794008), Sr. Manager for Procurement Lease Administration (data center/GPU/TPU leases) [E16](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[E12](https://job-boards.greenhouse.io/anthropic/jobs/5253835008), Head of FX & Risk [E32](https://job-boards.greenhouse.io/anthropic/jobs/5250433008), Real Estate Project Manager [E55](https://job-boards.greenhouse.io/anthropic/jobs/4939288008), Research Engineer for Code RL [E48](https://job-boards.greenhouse.io/anthropic/jobs/5254364008), and Applied AI Architect in Tokyo [E35](https://job-boards.greenhouse.io/anthropic/jobs/5076109008).\\n\\nGeographic distribution confirms multi-hub operations: San Francisco remains headquarters, with New York City, Seattle, London, Tokyo, Sydney, Ontario, Zürich, Boston, and Washington DC all represented [E20](https://job-boards.greenhouse.io/anthropic/jobs/5257650008)[E22](https://job-boards.greenhouse.io/anthropic/jobs/5255853008)[E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008)[E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008)[E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008)[E35](https://job-boards.greenhouse.io/anthropic/jobs/5076109008)[E58](https://job-boards.greenhouse.io/anthropic/jobs/5254803008)[E59](https://job-boards.greenhouse.io/anthropic/jobs/5252342008).\\n\\n## Data-business implications\\n\\n- **Data demand and labeling infrastructure**: The Human Data Platform PM role [E21](https://job-boards.greenhouse.io/anthropic/jobs/5195866008)[P15](https://job-boards.greenhouse.io/anthropic/jobs/5195866008) directly signals ongoing need for data labeling tooling, vendor interfaces, and data quality pipelines at scale. This implies spend on data vendors, crowd-worker platforms, and annotation infrastructure. The role's focus on \\\"novel interfaces for data vendors\\\" and \\\"pipelines that enable researchers to gather high-quality data at scale\\\" indicates that frontier model training continues to require substantial human-generated data, not just synthetic data.\\n\\n- **Evals ecosystem opportunity**: Claude Fable 5's benchmark claims—state-of-the-art on SWE-Bench Pro (+11 points), plus strength in knowledge work, vision, and scientific research [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/)—create demand for independent evaluation infrastructure. The tiered Mythos/Fable access model [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails) also creates a need for governance and compliance eval tooling that can verify model capability levels before granting access. The sleeper-agent probe work [P4](https://www.anthropic.com/research/probes-catch-sleeper-agents) suggests internal eval pipelines for detecting deceptive behavior are active and could inspire external safety eval products.\\n\\n- **Infrastructure tooling**: The inference runtime hiring [E20](https://job-boards.greenhouse.io/anthropic/jobs/5257650008)[P17](https://job-boards.greenhouse.io/anthropic/jobs/5257650008)[E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008)[E30](https://job-boards.greenhouse.io/anthropic/jobs/5152348008) explicitly mentions heterogeneous accelerator support (GPUs, TPUs, Trainium), signaling multi-vendor compute strategy. The TPM for API Platform role [E24](https://job-boards.greenhouse.io/anthropic/jobs/5256303008)[P13](https://job-boards.greenhouse.io/anthropic/jobs/5256303008) targets compute, database, networking, and inference coordination—indicating internal platform complexity that creates opportunities for infrastructure observability, orchestration, and cost-management tooling. The procurement lease administration role managing GPU/TPU leases [E16](https://github.com/anthropics/claude-code/releases/tag/v2.1.176)[E12](https://job-boards.greenhouse.io/anthropic/jobs/5253835008) confirms large-scale multi-vendor compute spend.\\n\\n- **Agent and deployment tooling**: The Claude Agent SDK shipping in both TypeScript and Python [P23](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176)[P25](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100)[P26](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177)[P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101), alongside the Claude Code Action for GitHub CI/CD [P24](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147), signals that Anthropic is building an agent runtime ecosystem. The 4.5% of public GitHub commits attributed to Claude Code [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code) suggests massive agent-driven code generation volume that creates demand for agent observability, cost tracking, and governance tooling. The background task lifecycle management in SDK v0.2.101 [P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101) indicates production-grade agent orchestration requirements.\\n\\n- **Safety and access-control products**: The Mythos/Fable tiered access model [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/)[W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails)[W6](https://www.youtube.com/watch?v=Y9Wz2PV404E) is effectively a safety-governance product: restricted model access for vetted partners via Project Glasswing. The Safeguards Rare Harms PM role [E50](https://job-boards.greenhouse.io/anthropic/jobs/5139628008) and GRC Platform Engineering Manager [E54](https://job-boards.greenhouse.io/anthropic/jobs/4980335008) suggest internal tooling for safety classification and compliance that could influence third-party safety infrastructure.\\n\\n- **Enterprise GTM and channel**: The TCS partnership [E5](https://www.anthropic.com/news/tcs-anthropic-partnership), DXC alliance [E53](https://www.anthropic.com/news/dxc-anthropic-alliance), Startup Partnerships manager [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008)[P18](https://job-boards.greenhouse.io/anthropic/jobs/5229558008), and System Integrator enablement lead [E52](https://job-boards.greenhouse.io/anthropic/jobs/5188391008) indicate a channel-first enterprise strategy. This creates integration, consulting, and managed-service opportunities in the Claude ecosystem, similar to early cloud platform buildouts.\\n\\n- **Product and GTM maturity**: The Field Reporting Insights Manager role [E51](https://job-boards.greenhouse.io/anthropic/jobs/5253257008)[P1](https://job-boards.greenhouse.io/anthropic/jobs/5253257008) specifying Salesforce, Looker, and BigQuery as the reporting stack signals a maturing revenue operations function. This implies data integration needs between CRM, product usage, and billing systems. The Web Producer and CMS Publishing role [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008)[P19](https://job-boards.greenhouse.io/anthropic/jobs/5257669008) indicates Anthropic is investing in its owned web surfaces (anthropic.com, claude.com) as product and brand channels.\\n\\nNo cited evidence supports direct vendor, revenue, or specific partnership revenue claims.\\n\\n## Traction highlights\\n\\n- Claude Fable 5/Mythos 5 launch: 2603 HN points / 2143 comments on the main launch post [E1](https://www.anthropic.com/news/claude-fable-5-mythos-5); 1306 points / 854 comments on the access-control follow-up [E2](https://www.anthropic.com/news/fable-mythos-access)—extraordinary developer and industry attention.\\n- Claude Code generates 4.5% of all public GitHub commits (~2.6 million weekly), and Anthropic engineers reportedly merge significantly more code when working with AI assistance [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code).\\n- Claude Fable 5 tops SWE-Bench Pro by 11 points over prior Claude models [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/); described by external observers as state-of-the-art on nearly all tested benchmarks [W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails).\\n- Fable 5 pricing at $10/$50 per million input/output tokens is less than half the price of the prior Mythos Preview tier [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/).\\n- GitHub Copilot integration: Fable 5 available in Copilot with enterprise admin policy controls [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/).\\n- Enterprise partner ecosystem forming: TCS [E5](https://www.anthropic.com/news/tcs-anthropic-partnership), DXC [E53](https://www.anthropic.com/news/dxc-anthropic-alliance), and system integrator channels [E52](https://job-boards.greenhouse.io/anthropic/jobs/5188391008) all cited.\\n- The Anthropic Institute launched May 2026 as a public research arm studying AI's real-world impacts [W5](https://www.anthropic.com/research/anthropic-institute-agenda).\\n- Interpretability research program has produced multiple high-profile publications across circuits, dictionary learning, and deception detection spanning 2021–2024 [P2](https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits)[P3](https://www.anthropic.com/research/in-context-learning-and-induction-heads)[P4](https://www.anthropic.com/research/probes-catch-sleeper-agents)[P5](https://www.anthropic.com/research/circuits-updates-april-2024)[P6](https://www.anthropic.com/research/circuits-updates-july-2024)[P7](https://www.anthropic.com/research/engineering-challenges-interpretability)[P8](https://www.anthropic.com/research/decomposing-language-models-into-understandable-components)[P10](https://www.anthropic.com/research/transformer-circuits)[P11](https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training)[P12](https://www.anthropic.com/research/circuits-updates-august-2024).\\n\\n## Sources\\n\\n- [P1](https://job-boards.greenhouse.io/anthropic/jobs/5253257008) Field Reporting Insights Manager job listing (Greenhouse)\\n- [P2](https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits) A Mathematical Framework For Transformer Circuits (Dec 2021)\\n- [P3](https://www.anthropic.com/research/in-context-learning-and-induction-heads) In-context Learning and Induction Heads (Mar 2022)\\n- [P4](https://www.anthropic.com/research/probes-catch-sleeper-agents) Probes Catch Sleeper Agents (Apr 2024)\\n- [P5](https://www.anthropic.com/research/circuits-updates-april-2024) Circuits Updates April 2024\\n- [P6](https://www.anthropic.com/research/circuits-updates-july-2024) Circuits Updates July 2024\\n- [P7](https://www.anthropic.com/research/engineering-challenges-interpretability) Engineering Challenges Interpretability (Jun 2024)\\n- [P8](https://www.anthropic.com/research/decomposing-language-models-into-understandable-components) Decomposing Language Models Into Understandable Components (Oct 2023)\\n- [P9](https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions) Evaluating And Mitigating Discrimination In Language Model Decisions (Dec 2023)\\n- [P10](https://www.anthropic.com/research/transformer-circuits) Transformer Circuits / Reflections on Qualitative Research (Mar 2024)\\n- [P11](https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training) Sleeper Agents Training Deceptive LLMs (Jan 2024)\\n- [P12](https://www.anthropic.com/research/circuits-updates-august-2024) Circuits Updates August 2024 (Sep 2024)\\n- [P13](https://job-boards.greenhouse.io/anthropic/jobs/5256303008) Technical Program Manager, API Platform job listing\\n- [P14](https://job-boards.greenhouse.io/anthropic/jobs/5255853008) IT Systems Engineer, Client Platform Engineer job listing\\n- [P15](https://job-boards.greenhouse.io/anthropic/jobs/5195866008) Product Management, Human Data Platform job listing\\n- [P16](https://job-boards.greenhouse.io/anthropic/jobs/5253835008) Sr. Manager, Procurement Lease Administration job listing\\n- [P17](https://job-boards.greenhouse.io/anthropic/jobs/5257650008) Staff+ Software Engineer, Inference Runtime job listing\\n- [P18](https://job-boards.greenhouse.io/anthropic/jobs/5229558008) Manager, Startup Partnerships job listing\\n- [P19](https://job-boards.greenhouse.io/anthropic/jobs/5257669008) Web Producer, CMS Publishing job listing\\n- [P20](https://job-boards.greenhouse.io/anthropic/jobs/5255912008) Engineering Manager, Enterprise job listing\\n- [P21](https://github.com/anthropics/claude-code/releases/tag/v2.1.177) anthropics/claude-code v2.1.177 release\\n- [P22](https://github.com/anthropics/claude-code/releases/tag/v2.1.176) anthropics/claude-code v2.1.176 release\\n- [P23](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176) anthropics/claude-agent-sdk-typescript v0.3.176 release\\n- [P24](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147) anthropics/claude-code-action v1.0.147 release\\n- [P25](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100) anthropics/claude-agent-sdk-python v0.2.100 release\\n- [P26](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177) anthropics/claude-agent-sdk-typescript v0.3.177 release\\n- [P27](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.148) anthropics/claude-code-action v1.0.148 release\\n- [P28](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101) anthropics/claude-agent-sdk-python v0.2.101 release\\n- [E1](https://www.anthropic.com/news/claude-fable-5-mythos-5) Claude Fable 5 Mythos 5 announcement\\n- [E2](https://www.anthropic.com/news/fable-mythos-access) Fable Mythos Access announcement\\n- [E3](https://www.anthropic.com/research/making-claude-a-chemist) Making Claude A Chemist\\n- [E4](https://www.anthropic.com/news/claude-corps) Claude Corps announcement\\n- [E5](https://www.anthropic.com/news/tcs-anthropic-partnership) TCS Anthropic Partnership\\n- [E6](https://www.anthropic.com/research/agents-in-biology) Agents In Biology\\n- [E7](https://job-boards.greenhouse.io/anthropic/jobs/5255912008) Engineering Manager, Enterprise job event\\n- [E8](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101) anthropics/claude-agent-sdk-python v0.2.101 release event\\n- [E9](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.148) anthropics/claude-code-action v1.0.148 release event\\n- [E10](https://github.com/anthropics/claude-code/releases/tag/v2.1.177) anthropics/claude-code v2.1.177 release event\\n- [E11](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177) anthropics/claude-agent-sdk-typescript v0.3.177 release event\\n- [E12](https://job-boards.greenhouse.io/anthropic/jobs/5253835008) Sr. Manager, Procurement Lease Administration job event\\n- [E13](https://job-boards.greenhouse.io/anthropic/jobs/5239794008) People Programs, M&A Lead job event\\n- [E14](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100) anthropics/claude-agent-sdk-python v0.2.100 release event\\n- [E15](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147) anthropics/claude-code-action v1.0.147 release event\\n- [E16](https://github.com/anthropics/claude-code/releases/tag/v2.1.176) anthropics/claude-code v2.1.176 release event\\n- [E17](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176) anthropics/claude-agent-sdk-typescript v0.3.176 release event\\n- [E18](https://job-boards.greenhouse.io/anthropic/jobs/5257669008) Web Producer, CMS Publishing job event\\n- [E19](https://job-boards.greenhouse.io/anthropic/jobs/5229558008) Manager, Startup Partnerships job event\\n- [E20](https://job-boards.greenhouse.io/anthropic/jobs/5257650008) Staff+ Software Engineer, Inference Runtime job event\\n- [E21](https://job-boards.greenhouse.io/anthropic/jobs/5195866008) Product Management, Human Data Platform job event\\n- [E22](https://job-boards.greenhouse.io/anthropic/jobs/5255853008) IT Systems Engineer, Client Platform Engineer job event\\n- [E23](https://job-boards.greenhouse.io/anthropic/jobs/5026187008) Field Marketing Lead, EMEA job event\\n- [E24](https://job-boards.greenhouse.io/anthropic/jobs/5256303008) Technical Program Manager, API Platform job event\\n- [E25](https://job-boards.greenhouse.io/anthropic/jobs/5109135008) Strategic Account Executive, Tech (SF) job event\\n- [E26](https://job-boards.greenhouse.io/anthropic/jobs/5255752008) Strategic Account Executive, Tech (Ontario) job event\\n- [E27](https://job-boards.greenhouse.io/anthropic/jobs/5198991008) Product Marketing Lead, Claude Platform - Cloud job event\\n- [E28](https://www.anthropic.com/news/anthropic-public-record) Anthropic Public Record announcement\\n- [E29](https://job-boards.greenhouse.io/anthropic/jobs/5097742008) Staff Software Engineer, Inference (London) job event\\n- [E30](https://job-boards.greenhouse.io/anthropic/jobs/5152348008) Sr. Software Engineer, Inference (London) job event\\n- [E31](https://job-boards.greenhouse.io/anthropic/jobs/5161882008) IT Systems Engineer, Enterprise SaaS job event\\n- [E32](https://job-boards.greenhouse.io/anthropic/jobs/5250433008) Head of FX & Risk job event\\n- [E33](https://job-boards.greenhouse.io/anthropic/jobs/5222180008) Account Executive - ASEAN job event\\n- [E34](https://job-boards.greenhouse.io/anthropic/jobs/4989228008) Manager, Customer Success (London) job event\\n- [E35](https://job-boards.greenhouse.io/anthropic/jobs/5076109008) Applied AI Architect (Tokyo) job event\\n- [E36](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.99) anthropics/claude-agent-sdk-python v0.2.99 release event\\n- [E37](https://job-boards.greenhouse.io/anthropic/jobs/5248946008) Strategy & Operations Lead, Marketing job event\\n- [E38](https://job-boards.greenhouse.io/anthropic/jobs/4802076008) IT Support Engineer (Seattle) job event\\n- [E39](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.146) anthropics/claude-code-action v1.0.146 release event\\n- [E40](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.175) anthropics/claude-agent-sdk-typescript v0.3.175 release event\\n- [E41](https://github.com/anthropics/claude-code/releases/tag/v2.1.175) anthropics/claude-code v2.1.175 release event\\n- [E42](https://job-boards.greenhouse.io/anthropic/jobs/5169167008) Field Marketing Manager job event\\n- [E43](https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.98) anthropics/claude-agent-sdk-python v0.2.98 release event\\n- [E44](https://github.com/anthropics/claude-code-action/releases/tag/v1.0.145) anthropics/claude-code-action v1.0.145 release event\\n- [E45](https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.174) anthropics/claude-agent-sdk-typescript v0.3.174 release event\\n- [E46](https://github.com/anthropics/claude-code/releases/tag/v2.1.174) anthropics/claude-code v2.1.174 release event\\n- [E47](https://job-boards.greenhouse.io/anthropic/jobs/4979585008) Product Support Specialist job event\\n- [E48](https://job-boards.greenhouse.io/anthropic/jobs/5254364008) Research Engineer, Code RL job event\\n- [E49](https://job-boards.greenhouse.io/anthropic/jobs/5254623008) Product Manager, GTM Experiences job event\\n- [E50](https://job-boards.greenhouse.io/anthropic/jobs/5139628008) Product Manager, Safeguards Rare Harms job event\\n- [E51](https://job-boards.greenhouse.io/anthropic/jobs/5253257008) Field Reporting Insights Manager job event\\n- [E52](https://job-boards.greenhouse.io/anthropic/jobs/5188391008) Partner Enablement Lead, System Integrators job event\\n- [E53](https://www.anthropic.com/news/dxc-anthropic-alliance) DXC Anthropic Alliance announcement\\n- [E54](https://job-boards.greenhouse.io/anthropic/jobs/4980335008) Engineering Manager, GRC Platform job event\\n- [E55](https://job-boards.greenhouse.io/anthropic/jobs/4939288008) Real Estate Project Manager job event\\n- [E56](https://job-boards.greenhouse.io/anthropic/jobs/5253339008) Product Manager, Enterprise job event\\n- [E57](https://job-boards.greenhouse.io/anthropic/jobs/5254582008) Policy Communications Manager job event\\n- [E58](https://job-boards.greenhouse.io/anthropic/jobs/5254803008) Staff+ Software Engineer, Developer Productivity job event\\n- [E59](https://job-boards.greenhouse.io/anthropic/jobs/5252342008) Senior/Staff Security Engineer, Threat Intelligence job event\\n- [E60](https://github.com/anthropics/leptos-chartistry) anthropics/leptos-chartistry fork event\\n- [W1](https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code) Claude writes 80% of its code, calls for AI pause (The Next Web, Jun 2026)\\n- [W2](https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/) Claude Fable 5 and the New AI Access-Control Race (Vastkind, Jun 2026)\\n- [W3](https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/) Claude Fable 5 Tops SWE-Bench Pro by 11 Points (GigaNectar, Jun 2026)\\n- [W4](https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails) Anthropic launches most powerful AI model yet (IBM, Jun 2026)\\n- [W5](https://www.anthropic.com/research/anthropic-institute-agenda) Focus areas for The Anthropic Institute (May 2026)\\n- [W6](https://www.youtube.com/watch?v=Y9Wz2PV404E) Introducing Claude Fable 5 (YouTube, Jun 2026)\",\n  \"cites\": [\n    \"P1\", \"P2\", \"P3\", \"P4\", \"P5\", \"P6\", \"P7\", \"P8\", \"P9\", \"P10\",\n    \"P11\", \"P12\", \"P13\", \"P14\", \"P15\", \"P16\", \"P17\", \"P18\", \"P19\", \"P20\",\n    \"P21\", \"P22\", \"P23\", \"P24\", \"P25\", \"P26\", \"P27\", \"P28\",\n    \"E1\", \"E2\", \"E3\", \"E4\", \"E5\", \"E6\", \"E7\", \"E8\", \"E9\", \"E10\",\n    \"E11\", \"E12\", \"E13\", \"E14\", \"E15\", \"E16\", \"E17\", \"E18\", \"E19\", \"E20\",\n    \"E21\", \"E22\", \"E23\", \"E24\", \"E25\", \"E26\", \"E27\", \"E28\", \"E29\", \"E30\",\n    \"E31\", \"E32\", \"E33\", \"E34\", \"E35\", \"E36\", \"E37\", \"E38\", \"E39\", \"E40\",\n    \"E41\", \"E42\", \"E43\", \"E44\", \"E45\", \"E46\", \"E47\", \"E48\", \"E49\", \"E50\",\n    \"E51\", \"E52\", \"E53\", \"E54\", \"E55\", \"E56\", \"E57\", \"E58\", \"E59\", \"E60\",\n    \"W1\", \"W2\", \"W3\", \"W4\", \"W5\", \"W6\"\n  ]\n}\n```","generated_at":"2026-06-13T08:04:04.953+00:00","citations":[{"url":"https://www.vastkind.com/claude-fable-5-frontier-ai-access-control/","path":null,"label":"vastkind.com/claude-fable-5-frontier-ai-access-control","type":"external"},{"url":"https://www.ibm.com/think/news/anthropic-launches-powerful-ai-model-fable-5-mythos-5-safety-guardrails","path":null,"label":"ibm.com/think","type":"external"},{"url":"https://www.youtube.com/watch?v=Y9Wz2PV404E","path":null,"label":"youtube.com/watch","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5026187008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5109135008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5255752008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5222180008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/4989228008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5076109008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.177","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.176","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code","path":null,"label":"thenextweb.com/news","type":"external"},{"url":"https://www.anthropic.com/research/probes-catch-sleeper-agents","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/circuits-updates-april-2024","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/circuits-updates-july-2024","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/engineering-challenges-interpretability","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/circuits-updates-august-2024","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://github.com/anthropics/leptos-chartistry","path":null,"label":"anthropics/leptos-chartistry","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5255912008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5255912008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5253339008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5229558008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5229558008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5188391008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5198991008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5169167008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5257650008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5257650008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5097742008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5152348008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5256303008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5256303008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5195866008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5195866008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5253257008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5253257008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5139628008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5254582008","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/5252342008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5254803008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5255853008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5255853008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5161882008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5239794008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.176","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5253835008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.177","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.175","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.174","path":null,"label":"anthropics/claude-code","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.177","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.176","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.175","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-typescript/releases/tag/v0.3.174","path":null,"label":"anthropics/claude-agent-sdk-typescript","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.101","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.100","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.99","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-agent-sdk-python/releases/tag/v0.2.98","path":null,"label":"anthropics/claude-agent-sdk-python","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.148","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.145","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.147","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://www.anthropic.com/news/claude-fable-5-mythos-5","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026/","path":null,"label":"giganectar.com/claude-fable-5-mythos-5-benchmarks-pricing-safeguards-anthropic-2026","type":"external"},{"url":"https://www.anthropic.com/news/fable-mythos-access","path":null,"label":"anthropic.com/news","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/research/agents-in-biology","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/evaluating-and-mitigating-discrimination-in-language-model-decisions","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/transformer-circuits","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/news/tcs-anthropic-partnership","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/dxc-anthropic-alliance","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/news/claude-corps","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/news/anthropic-public-record","path":null,"label":"anthropic.com/news","type":"external"},{"url":"https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/in-context-learning-and-induction-heads","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/decomposing-language-models-into-understandable-components","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training","path":null,"label":"anthropic.com/research","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5254364008","path":null,"label":"job-boards.greenhouse.io/anthropic","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/5254623008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/4979585008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5257669008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5250433008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/4939288008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5257669008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/5253835008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.148","path":null,"label":"anthropics/claude-code-action","type":"external"},{"url":"https://job-boards.greenhouse.io/anthropic/jobs/4802076008","path":null,"label":"job-boards.greenhouse.io/anthropic","type":"external"},{"url":"https://github.com/anthropics/claude-code-action/releases/tag/v1.0.146","path":null,"label":"anthropics/claude-code-action","type":"external"}],"provenance":{"provider":"deepseek","model":"deepseek-v4-pro","workflow":"onlylabs-deepagents-analysis-v3","agent":"deepagents"},"evidence":{"total":94,"pages":28,"events":140,"web":6,"signal_desks":{"forks":1,"repos":0,"hiring":35,"talking":8,"releases":16},"data_radar_lanes":{"data":1,"evals":0,"safety":3,"product":11,"infrastructure":9},"data_radar_matches":20}},{"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":"{\"content\": \"## Thesis\\n\\nAmazon is operating on two reinforcing tracks: hardening its infrastructure moat through silicon, networking, and formal verification while simultaneously standing up an integrated agentic AI platform. The lab is not competing on raw parameter count — it is publicly arguing that \\\"intelligence isn't about parameter count. It's about time\\\" — and instead betting on inference-time compute,…","markdown":"{\"content\": \"## Thesis\\n\\nAmazon is operating on two reinforcing tracks: hardening its infrastructure moat through silicon, networking, and formal verification while simultaneously standing up an integrated agentic AI platform. The lab is not competing on raw parameter count — it is publicly arguing that \\\"intelligence isn't about parameter count. It's about time\\\" [E41](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time) — and instead betting on inference-time compute, RL-trained agents, and deep vertical integration from custom silicon through software to deployment. Graviton5's chiplet architecture explicitly targets \\\"agentic AI workloads\\\" [P4](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law)[E2](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law), the Nitro Isolation Engine provides mathematical assurance of VM isolation via formal verification [P1](https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/compliance/nitro-isolation-engine-whitepaper.pdf)[P5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation), and data center network topology research pursues flatter, more efficient designs [E3](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks). On the AI side, Nova Act trains \\\"model capabilities, orchestration logic, and tool controls together as one integrated system\\\" [W3](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents), and the Amazon AGI Labs are building a perception agent harness in the open [W2](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source). The publicly visible research portfolio is eval-heavy, agent-centric, and safety-conscious — signaling a lab building trustworthy production AI for enterprise deployment rather than chasing benchmark leaderboards.\\n\\n## Signal desks\\n\\n### Hiring\\n\\n- **Organizational structure confirmed but no job postings cited**: Amazon AGI is the model-building org, owning the Olympus training program and the Nova model family, with research staff reporting through Rohit Prasad. AWS AI Services sits inside the cloud business and ships Bedrock, SageMaker, Q, and Comprehend [W4](https://ctaio.dev/en/salary/amazon-agi-salary/). This bifurcation implies hiring demand across both research (AGI Labs) and product-engineering (AWS AI Services), but no specific roles, locations, team expansions, or job descriptions can be confirmed from this evidence pack.\\n- **Implied clusters from research output**: The breadth of open-source repos and publications suggests active staffing in agentic AI, evals/benchmarks, model efficiency, multimodal, code/SWE, safety, and robotics. However, this is inferential only.\\n- **Gap**: No LinkedIn posts, job board listings, or careers-page evidence supplied. Hiring remains a blind spot in this pack.\\n\\n### Forks\\n\\n- **No cited evidence in this pack.** All repositories surfaced are original `amazon-science` repos, not forks of upstream projects. The MXFP4-LLM repo references and depends on upstream projects including NVIDIA/Megatron-LM, NVIDIA/TransformerEngine, and microsoft/microxcaling [P9](https://github.com/amazon-science/mxfp4-llm), but no fork events were captured. The absence of fork activity in the evidence may reflect Amazon's preference for original implementation over upstream contribution through forking, but no conclusion can be drawn from the data provided.\\n\\n### Releases\\n\\n- **HQwen3 \\\"primed\\\" family (March 2026)**: Amazon released at least 14 fine-tuned model variants on HuggingFace under Apache-2.0 license, centered on HQwen3 architectures at 8B and 32B scales — including GKA-primed, GDN-primed, BMOJOF-primed, and Mamba2-primed variants across both Instruct and Reasoner configurations [E43](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct)[E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)[E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner)[E48](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner)[E49](https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct)[E51](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct)[E52](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct)[E53](https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct)[E54](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner). The GKA-primed-HQwen3-32B-Instruct leads with 46,777 downloads [E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct). This signals a systematic fine-tuning pipeline producing distinct model variants for different deployment profiles.\\n- **GPT-OSS EAGLE long-context family (Feb–May 2026)**: 20B, 120B, and Qwen3-Coder-30B variants with speculative decoding (EAGLE) and long-context support [E15](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E17](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E18](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context)[E57](https://huggingface.co/amazon/gpt-oss-120b-p-eagle). Apache-2.0 licensed; modest traction so far (48–254 downloads).\\n- **Chronos-2 (Oct 2025)**: Time-series forecasting model with 12.5M HuggingFace downloads, 322 likes, 119M parameters, Apache-2.0 [E1](https://huggingface.co/amazon/chronos-2). The most-downloaded Amazon model in this evidence set — a utility model with clear product-market fit distinct from the LLM line.\\n- **Tooling releases**: UniqSketch v1.3.0 — genomic Bloom filter sizing with auto-calibration features [P2](https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0)[E4](https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0); Concurry v0.13.x — Python concurrency library for AI workloads [E13](https://github.com/amazon-science/concurry/releases/tag/v0.13.2)[E14](https://github.com/amazon-science/concurry/releases/tag/v0.13.1); azcausal v0.2.4–v0.2.5 — causal inference framework [E33](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E56](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3).\\n- **Pattern**: The release cadence shows heavy investment in model fine-tuning infrastructure (the \\\"primed\\\" series), speculative decoding for inference acceleration (EAGLE variants), and domain-specific tooling (genomics, causal inference, parallelism).\\n\\n### Talking\\n\\nAmazon's public communications cluster into five major themes:\\n\\n- **Agentic AI as product surface**: Nova Act is framed as an agent-building service that \\\"trains model capabilities, orchestration logic, and tool controls together\\\" to address the trust gap keeping agents experimental [W3](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents). The perception agent harness with annotation and verification primitives is being built in the open on GitHub [W2](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source). Posts on \\\"real-world grounding in agentic AI\\\" [E6](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai), \\\"bridging intent and execution in agentic systems\\\" [E7](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems), and RuleForge's agentic vulnerability detection achieving 336% faster detection rules [E45](https://www.amazon.science/blog/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale) reinforce the agent narrative.\\n- **Infrastructure as differentiator**: Graviton5's chiplet architecture delivers 25% better performance for \\\"general-purpose and agentic AI workloads\\\" with DDR5-8800 and PCIe gen6 [P4](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law)[E2](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law). \\\"Quasi-random\\\" flat network topologies and ShuffleBox optical components promise more efficient data center fabrics [E3](https://www.amazon.science/blog/how-flat-is-replacing-fat-in-aws-data-center-networks). The Nitro Isolation Engine is positioned as the \\\"first formally verified cloud hypervisor\\\" [P5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation), with Isabelle/HOL enabling the proof [E36](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine).\\n- **Evals, data, and ground truth**: \\\"Ground truth is a process, not a dataset\\\" [E9](https://www.amazon.science/blog/ground-truth-is-a-process-not-a-dataset) directly addresses the benchmarking challenge for long-form AI outputs. The Antibody Developability Benchmark with Johns Hopkins is \\\"one of the most diverse antibody datasets in public literature\\\" [E40](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design). Amazon Research Awards provide recipients access to \\\"Amazon public datasets, along with AWS AI/ML services and tools\\\" across 49 universities in 11 countries [E11](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced).\\n- **Safety and trust**: A dedicated post on \\\"building trust into AI\\\" describes the responsible-AI pipeline embedding \\\"safety and values throughout the AI development lifecycle\\\" [E31](https://www.amazon.science/blog/building-trust-into-ai). Formal verification of the Nitro hypervisor using a restricted Rust subset [P1](https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/compliance/nitro-isolation-engine-whitepaper.pdf)[E5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E36](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine), post-quantum cryptography deployment with automated reasoning [E46](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon), training data privacy attacks and cryptographic defenses [E34](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data), and a statistical framework for estimating catastrophic LLM failure likelihood [E35](https://www.amazon.science/blog/how-catastrophic-is-your-llm) form a comprehensive safety narrative.\\n- **Efficiency over scale**: \\\"Intelligence isn't about parameter count. It's about time\\\" argues for reducing inference time as models grow [E41](https://www.amazon.science/blog/intelligence-isnt-about-parameter-count-its-about-time). A scaling law post claims architectural choices improve throughput by \\\"up to 47% with no loss of accuracy\\\" [E20](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy). The MXFP4 training recipe uses stochastic rounding and random Hadamard transforms for near-lossless 4-bit training [P9](https://github.com/amazon-science/mxfp4-llm).\\n\\n## Shipping\\n\\nAmazon shipped materially across four vectors in the evidence window:\\n\\n1. **Models**: Chronos-2 (time-series, 119M params, 12.5M+ downloads) [E1](https://huggingface.co/amazon/chronos-2); HQwen3-primed Instruct and Reasoner family at 8B and 32B scales, Apache-2.0 [E43](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct)[E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)[E47](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Reasoner)[E48](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Reasoner)[E49](https://huggingface.co/amazon/GDN-primed-HQwen3-32B-Instruct)[E51](https://huggingface.co/amazon/GKA-primed-HQwen3-8B-Instruct)[E52](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Instruct)[E53](https://huggingface.co/amazon/BMOJOF-primed-HQwen3-8B-Instruct)[E54](https://huggingface.co/amazon/GDN-primed-HQwen3-8B-Reasoner); GPT-OSS EAGLE long-context variants at 20B, 120B, and Qwen3-Coder-30B [E15](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E17](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E18](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context)[E57](https://huggingface.co/amazon/gpt-oss-120b-p-eagle); Mamba2-primed hybrid SSM-attention architecture [E43](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct).\\n2. **Infrastructure**: Graviton5 silicon with chiplet architecture and DDR5-8800/PCIe gen6 interconnects [P4](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law)[E2](https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law); Nitro Isolation Engine with formal verification proofs published as whitepaper [P1](https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/compliance/nitro-isolation-engine-whitepaper.pdf)[P5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation).\\n3. **Tooling**: Concurry parallelism library v0.13.x on PyPI [P6](https://github.com/amazon-science/concurry)[E13](https://github.com/amazon-science/concurry/releases/tag/v0.13.2)[E14](https://github.com/amazon-science/concurry/releases/tag/v0.13.1); UniqSketch v1.3.0 with auto-calibration [P2](https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0)[E4](https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0); azcausal v0.2.5 [E33](https://github.com/amazon-science/azcausal/releases/tag/v0.2.5)[E56](https://github.com/amazon-science/azcausal/releases/tag/v0.2.4.3); Nova Act Skills SDK and Annotator Browser Extension on GitHub [W2](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source).\\n4. **Platform**: Amazon Nova Forge for custom model training with hyperparameter optimization guidance, including data mixing to \\\"blend your training data with curated datasets\\\" and prevent catastrophic forgetting [W1](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/); Nova Act agent-building service [W3](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents).\\n\\n## Research themes\\n\\n**Agentic AI dominates the portfolio.** The single largest research cluster spans reskill (agent RL training with skill co-evolution, built as a veRL extension) [E8](https://github.com/amazon-science/reskill), DualKV (shared-prompt Flash Attention for efficient RL training with large rollouts and long contexts) [E10](https://github.com/amazon-science/dualkv-flash-attn-for-rl), EvoMAS (evolutionary generation of multi-agent systems, ICML 2026) [E16](https://github.com/amazon-science/EvoMAS), PROF-GRPO [E24](https://github.com/amazon-science/PROF-GRPO), agentic-forking-path [E60](https://github.com/amazon-science/agentic-forking-path), compagent (visual compliance verification) [E28](https://github.com/amazon-science/compagent), and QualityFlow (agentic program synthesis with LLM Quality Checker, SOTA on MBPP and HumanEval) [P28](https://github.com/amazon-science/QualityFlow). This is reinforced by public writing on agent reliability, grounding, and the Nova Act perception harness [W2](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source)[W3](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents)[E6](https://www.amazon.science/blog/real-world-grounding-in-agentic-ai)[E7](https://www.amazon.science/blog/bridging-intent-and-execution-in-agentic-systems).\\n\\n**Evals and benchmarks form a second pillar.** Document Haystack (long-context multimodal VLM benchmark, 400 document variants, 8,250 questions) [P8](https://github.com/amazon-science/document-haystack); GaRAGe (2,366 RAG questions with 35K+ annotated grounding passages, ACL 2025 Findings) [P19](https://github.com/amazon-science/GaRAGe); CiteEval (principle-driven citation evaluation framework with CiteBench) [P27](https://github.com/amazon-science/CiteEval); MigrationBench (code migration evaluation framework with Java and Python support) [P13](https://github.com/amazon-science/MigrationBench)[P10](https://github.com/amazon-science/JavaMigration); PersonaLens (personalization benchmark for conversational AI, ACL 2025) [P17](https://github.com/amazon-science/PersonaLens); ConFETTI (conversational function-calling evaluation, 109 conversations, 313 user turns, 86 APIs) [P16](https://github.com/amazon-science/confetti); TN-Eval (behavioral therapy note quality rubric, ACL 2025 Industry) [P23](https://github.com/amazon-science/TN-Eval)[P24](https://github.com/amazon-science/TN-Eval-Data); RMIR (reasoning-intensive multimodal image retrieval benchmark) [E32](https://github.com/amazon-science/rmir); ACI-bench hallucination annotations with expert-labeled severity categories [P14](https://github.com/amazon-science/acibench-hallucination-annotations); TrivialPlus (long-context hallucination detection benchmark, ACL 2026 main) [E25](https://github.com/amazon-science/hallucination-benchmark-trivialplus); temporal reasoning dataset for multilingual temporal reasoning [E23](https://github.com/amazon-science/temporal-reasoning-dataset); TISER (timeline self-reflection for temporal reasoning, ACL 2025) [P18](https://github.com/amazon-science/TISER); Query-Conditioned NLI [P20](https://github.com/amazon-science/Query-Conditioned-NLI); and XRAG (cross-lingual retrieval-augmented generation) [P15](https://github.com/amazon-science/XRAG).\\n\\n**Efficiency, compression, and scaling**: MXFP4 training recipe achieving near-lossless training via unbiased gradient estimates with stochastic rounding and random Hadamard transforms [P9](https://github.com/amazon-science/mxfp4-llm); ProxSparse (regularized learning of 2:4 semi-structured sparsity masks, ICML 2025) [P26](https://github.com/amazon-science/ProxSparse); scaling laws for architectural choices yielding up to 47% throughput improvement without accuracy loss [E20](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy); expert upcycling [E37](https://github.com/amazon-science/expert-upcycling); adaptive layerwise perturbation [E21](https://github.com/amazon-science/adaptive-layerwise-perturbation); information preservation in prompt compression (EMNLP 2025) [P12](https://github.com/amazon-science/information-preservation-in-prompt-compression); and Promptimus automated prompt engineering targeting specific failure points [E22](https://www.amazon.science/blog/promptimus-improving-already-good-llm-prompts-with-zero-manual-engineering).\\n\\n**Robotics and embodied AI**: Spherical Diffusion Policy (SE(3) equivariant visuomotor policy, ICML 2025, benchmarked on 20 MimicGen simulation tasks and 5 physical robot tasks) [P25](https://github.com/amazon-science/Spherical_Diffusion_Policy); TransitionFlowMatching (image and video generation via transition matching, AISTATS 2026) [E42](https://github.com/amazon-science/TransitionFlowMatching).\\n\\n**Additional active threads**: diverse reasoning traces using tokens to control distinct reasoning strategies [E12](https://www.amazon.science/blog/diverse-reasoning-traces-teach-llms-to-make-better-decisions); LLM-based text-to-speech with LoRA, data augmentation, and chain-of-thought reasoning [E58](https://www.amazon.science/blog/improving-quality-and-robustness-in-llm-based-text-to-speech-systems); audio retrieval with complex text queries (LARCQ, Interspeech 2025) [P22](https://github.com/amazon-science/LARCQ); OmniMatch for joinability discovery in data products [P21](https://github.com/amazon-science/omnimatch); SWAN semantic watermarking with abstract meaning representation (ACL 2026) [E29](https://github.com/amazon-science/SWAN); JavaMigration LLM-based code migration agent built on Strands Agents [P10](https://github.com/amazon-science/JavaMigration); customized Nova models for molecular-property prediction in drug discovery [E39](https://www.amazon.science/blog/customized-amazon-nova-models-improve-molecular-property-prediction-in-drug-discovery); mechanism design theory applied to Amazon-vendor supply chain collaboration [E30](https://www.amazon.science/blog/how-mechanism-design-theory-helps-optimize-amazon-vendor-collaboration); and middle-mile delivery network optimization under uncertainty [E26](https://www.amazon.science/blog/navigating-uncertainty-in-amazons-middle-mile-network).\\n\\n## Hiring & scaling\\n\\n**No cited job postings in this evidence pack.** The only organizational signal is [W4](https://ctaio.dev/en/salary/amazon-agi-salary/), which establishes that Amazon AGI (led by Rohit Prasad) is the model-building organization responsible for the Olympus training program and Nova model family, while AWS AI Services ships Bedrock, SageMaker, Q, and Comprehend. This split creates hiring demand across both research (AGI Labs) and product-engineering (AWS AI Services) functions. The breadth of the public research portfolio — spanning agents, evals, efficiency, robotics, multimodal, audio, code, safety, and drug discovery — implies a large, distributed research staff, but specific headcount, growth rate, open roles, locations, or team expansions cannot be confirmed from this pack.\\n\\n## Data-business implications\\n\\n**Evals infrastructure demand**: Amazon is producing evaluation artifacts at high velocity — Document Haystack [P8](https://github.com/amazon-science/document-haystack), GaRAGe [P19](https://github.com/amazon-science/GaRAGe), CiteEval [P27](https://github.com/amazon-science/CiteEval), MigrationBench [P13](https://github.com/amazon-science/MigrationBench), PersonaLens [P17](https://github.com/amazon-science/PersonaLens), ConFETTI [P16](https://github.com/amazon-science/confetti), TN-Eval [P23](https://github.com/amazon-science/TN-Eval), RMIR [E32](https://github.com/amazon-science/rmir), and multiple hallucination benchmarks [P14](https://github.com/amazon-science/acibench-hallucination-annotations)[E25](https://github.com/amazon-science/hallucination-benchmark-trivialplus). These require hosting, scoring infrastructure, and LLM-as-judge pipelines. The \\\"ground truth is a process, not a dataset\\\" framing [E9](https://www.amazon.science/blog/ground-truth-is-a-process-not-a-dataset) implies ongoing annotation and re-annotation workflows, creating sustained demand for human-and-model-in-the-loop eval tooling.\\n\\n**Agent infrastructure and orchestration**: Nova Act's architecture — training model capabilities, orchestration logic, and tool controls together as one integrated system [W3](https://www.aboutamazon.com/news/aws/how-amazon-builds-ai-agents) — plus the perception agent harness with annotation and verification primitives [W2](https://labs.amazon.science/blog/introducing-the-perception-agent-harness-annotation-and-verification-open-source), reskill for RL-based agent training [E8](https://github.com/amazon-science/reskill), and DualKV for efficient RL rollouts [E10](https://github.com/amazon-science/dualkv-flash-attn-for-rl) signal investment in agent-specific training and inference infrastructure distinct from standard LLM serving. This creates opportunities in RL training frameworks, agent evaluation harnesses, and orchestration middleware.\\n\\n**Data and dataset curation as moat**: The Nova Forge hyperparameter post emphasizes data mixing to blend custom training data with curated datasets to prevent catastrophic forgetting [W1](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/). The Antibody Developability Benchmark is explicitly \\\"powered by one of the most diverse antibody datasets in public literature\\\" [E40](https://www.amazon.science/news/aws-gray-lab-johns-hopkins-announce-groundbreaking-database-for-ai-ml-antibody-design). Amazon Research Awards provide academic access to \\\"Amazon public datasets\\\" [E11](https://www.amazon.science/research-awards/latest-news/fall-2025-amazon-research-awards-recipients-announced). Privacy-preserving training research reproduces and defends against three data extraction attacks with cryptographic defenses [E34](https://www.amazon.science/blog/preserving-the-privacy-of-ai-training-data). These signals point to strategic data-asset construction and data-governance infrastructure investment.\\n\\n**Safety and security as product differentiator**: Formal verification of the Nitro hypervisor using Isabelle/HOL and a restricted Rust subset [P1](https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/compliance/nitro-isolation-engine-whitepaper.pdf)[P5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E5](https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation)[E36](https://www.amazon.science/blog/isabelle-hol-the-proof-assistant-behind-the-nitro-isolation-engine) is a cloud-security product claim, not pure research. Post-quantum cryptography deployment reconciling \\\"security, performance, and maintainability\\\" [E46](https://www.amazon.science/blog/verifying-and-optimizing-post-quantum-cryptography-at-amazon) and the responsible-AI pipeline embedding \\\"safety and values throughout the AI development lifecycle\\\" [E31](https://www.amazon.science/blog/building-trust-into-ai) are enterprise-trust signals. TurboFuzzLLM for automated LLM red-teaming [P11](https://github.com/amazon-science/TurboFuzzLLM) and the catastrophic-failure estimation framework for adversarial conversations [E35](https://www.amazon.science/blog/how-catastrophic-is-your-llm) address enterprise compliance and safety requirements.\\n\\n**Inference optimization for deployment**: The EAGLE long-context speculative decoding releases [E15](https://huggingface.co/amazon/gpt-oss-20b-p-eagle-long-context)[E17](https://huggingface.co/amazon/Qwen3-Coder-30B-A3B-Instruct-P-EAGLE-long-context)[E18](https://huggingface.co/amazon/gpt-oss-120b-p-eagle-long-context), throughput-optimized scaling laws claiming 47% improvement [E20](https://www.amazon.science/blog/making-llms-faster-without-sacrificing-accuracy), MXFP4 training for 4-bit compute [P9](https://github.com/amazon-science/mxfp4-llm), ProxSparse 2:4 structured sparsity [P26](https://github.com/amazon-science/ProxSparse), and Mamba2-primed hybrid architecture exploration [E43](https://huggingface.co/amazon/Mamba2-primed-HQwen3-8B-Instruct) all target deployment efficiency — reducing cost-per-token for production workloads.\\n\\n**Product market signals**: Chronos-2 at 12.5M HuggingFace downloads [E1](https://huggingface.co/amazon/chronos-2) demonstrates product-market fit for open time-series models. The HQwen3-primed family's consistent Apache-2.0 licensing [E43–E54] lowers enterprise adoption friction. Amazon Nova Forge's detailed hyperparameter optimization guidance [W1](https://aws.amazon.com/blogs/machine-learning/the-art-and-science-of-hyperparameter-optimization-on-amazon-nova-forge/) directly supports the enterprise fine-tuning customer pipeline on AWS.\\n\\n## Traction highlights\\n\\n- **Chronos-2**: 12.5M+ HuggingFace downloads, 322 likes, Apache-2.0 — the most-downloaded Amazon model in this evidence set [E1](https://huggingface.co/amazon/chronos-2)\\n- **GKA-primed-HQwen3-32B-Instruct**: 46,777 downloads, leading the \\\"primed\\\" family [E44](https://huggingface.co/amazon/GKA-primed-HQwen3-32B-Instruct)\\n- **MXFP4-LLM**: 127 GitHub stars, 18 forks [P9](https://github.com/amazon-science/mxfp4-llm)\\n- **JuLS**: 176 GitHub stars, 6 HN points/4 comments [E19](https://github.com/amazon-science/JuLS)\\n- **Spherical Diffusion Policy**: 43 GitHub stars, ICML 2025 [P25](https://github.com/amazon-science/Spherical_Diffusion_Policy)\\n- **TurboFuzzLLM**: 24 GitHub stars [P11](https://github.com/amazon-science/TurboFuzzLLM)\\n- **Concurry**: 18 GitHub stars, PyPI-published [P6](https://github.com/amazon-science/concurry)\\n- **MigrationBench**: 14 GitHub stars [P13](https://github.com/amazon-science/MigrationBench)\\n- **GaRAGe**: 13 GitHub stars, ACL 2025 Findings [P19](https://github.com/amazon-science/GaRAGe)\\n- **TISER**: 13 GitHub stars, ACL 2025 Main [P18](https://github.com/amazon-science/TISER)\\n- **Multiple top-tier acceptances**: ACL 2025 (PersonaLens, TISER, GaRAGe, TN-Eval) [P17](https://github.com/amazon-science/PersonaLens)[P18](https://github.com/amazon-science/TISER)[P19](https://github.com/amazon-science/GaRAGe)[P23](https://github.com/amazon-science/TN-Eval); ICML 2025/2026 (ProxSparse, Spherical Diffusion Policy, EvoMAS) [P25](https://github.com/amazon-science/Spherical_Diffusion_Policy)[P26](https://github.com/amazon-science/ProxSparse)[E16](https://github.com/amazon-science/EvoMAS); EMNLP 2025 (prompt compression) [P12](https://github.com/amazon-science/information-preservation-in-prompt-compression); Interspeech 2025 (LARCQ) [P22](https://github.com/amazon-science/LARCQ); AISTATS 2026 (TransitionFlowMatching) [E42](https://github.com/amazon-science/TransitionFlowMatching); ACL 2026 (TrivialPlus, SWAN) [E25](https://github.com/amazon-science/hallucination-benchmark-trivialplus)[E29](https://github.com/amazon-science/SWAN)\\n\\n## Sources\\n\\nEvidence pack spanning pages [P1–P28], events [E1–E60], and web captures [W1–W4].\", \"cites\": [\"P1\", \"P2\", \"P4\", \"P5\", \"P6\", \"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\", \"E28\", \"E29\", \"E30\", \"E31\", \"E32\", \"E33\", \"E34\", \"E35\", \"E36\", \"E37\", \"E39\", \"E40\", \"E41\", \"E42\", \"E43\", \"E44\", \"E45\", \"E46\", \"E47\", \"E48\", \"E49\", \"E51\", \"E52\", \"E53\", \"E54\", \"E56\", \"E57\", \"E58\", \"E60\", \"W1\", \"W2\", \"W3\", \"W4\"]}</｜｜DSML｜｜parameter>\n</｜｜DSML｜｜invoke>\n</｜｜DSML｜｜tool_calls>","generated_at":"2026-06-13T08:02:30.785+00:00","citations":[{"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/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/graviton5s-improved-design-increases-speed-and-energy-efficiency-beyond-moores-law","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://d1.awsstatic.com/onedam/marketing-channels/website/aws/en_US/whitepapers/compliance/nitro-isolation-engine-whitepaper.pdf","path":null,"label":"d1.awsstatic.com/onedam","type":"external"},{"url":"https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://www.amazon.science/blog/ec2s-formally-verified-isolation-engine-provides-mathematical-assurance-of-virtual-machine-isolation","path":null,"label":"amazon.science/blog","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.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://ctaio.dev/en/salary/amazon-agi-salary/","path":null,"label":"ctaio.dev/en","type":"external"},{"url":"https://github.com/amazon-science/mxfp4-llm","path":null,"label":"amazon-science/mxfp4-llm","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-HQwen3-32B-Instruct","path":null,"label":"amazon/GKA-primed-HQwen3-32B-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://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/chronos-2","path":null,"label":"amazon/chronos-2","type":"external"},{"url":"https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0","path":null,"label":"amazon-science/uniqsketch","type":"external"},{"url":"https://github.com/amazon-science/uniqsketch/releases/tag/v1.3.0","path":null,"label":"amazon-science/uniqsketch","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://www.amazon.science/blog/real-world-grounding-in-agentic-ai","path":null,"label":"amazon.science/blog","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/how-amazon-uses-agentic-ai-for-vulnerability-detection-at-global-scale","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/ground-truth-is-a-process-not-a-dataset","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/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/blog/building-trust-into-ai","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/how-catastrophic-is-your-llm","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://github.com/amazon-science/concurry","path":null,"label":"amazon-science/concurry","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://github.com/amazon-science/reskill","path":null,"label":"amazon-science/reskill","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/EvoMAS","path":null,"label":"amazon-science/EvoMAS","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/agentic-forking-path","path":null,"label":"amazon-science/agentic-forking-path","type":"external"},{"url":"https://github.com/amazon-science/compagent","path":null,"label":"amazon-science/compagent","type":"external"},{"url":"https://github.com/amazon-science/QualityFlow","path":null,"label":"amazon-science/QualityFlow","type":"external"},{"url":"https://github.com/amazon-science/document-haystack","path":null,"label":"amazon-science/document-haystack","type":"external"},{"url":"https://github.com/amazon-science/GaRAGe","path":null,"label":"amazon-science/GaRAGe","type":"external"},{"url":"https://github.com/amazon-science/CiteEval","path":null,"label":"amazon-science/CiteEval","type":"external"},{"url":"https://github.com/amazon-science/MigrationBench","path":null,"label":"amazon-science/MigrationBench","type":"external"},{"url":"https://github.com/amazon-science/JavaMigration","path":null,"label":"amazon-science/JavaMigration","type":"external"},{"url":"https://github.com/amazon-science/PersonaLens","path":null,"label":"amazon-science/PersonaLens","type":"external"},{"url":"https://github.com/amazon-science/confetti","path":null,"label":"amazon-science/confetti","type":"external"},{"url":"https://github.com/amazon-science/TN-Eval","path":null,"label":"amazon-science/TN-Eval","type":"external"},{"url":"https://github.com/amazon-science/TN-Eval-Data","path":null,"label":"amazon-science/TN-Eval-Data","type":"external"},{"url":"https://github.com/amazon-science/rmir","path":null,"label":"amazon-science/rmir","type":"external"},{"url":"https://github.com/amazon-science/acibench-hallucination-annotations","path":null,"label":"amazon-science/acibench-hallucination-annotations","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/temporal-reasoning-dataset","path":null,"label":"amazon-science/temporal-reasoning-dataset","type":"external"},{"url":"https://github.com/amazon-science/TISER","path":null,"label":"amazon-science/TISER","type":"external"},{"url":"https://github.com/amazon-science/Query-Conditioned-NLI","path":null,"label":"amazon-science/Query-Conditioned-NLI","type":"external"},{"url":"https://github.com/amazon-science/XRAG","path":null,"label":"amazon-science/XRAG","type":"external"},{"url":"https://github.com/amazon-science/ProxSparse","path":null,"label":"amazon-science/ProxSparse","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/adaptive-layerwise-perturbation","path":null,"label":"amazon-science/adaptive-layerwise-perturbation","type":"external"},{"url":"https://github.com/amazon-science/information-preservation-in-prompt-compression","path":null,"label":"amazon-science/information-preservation-in-prompt-compression","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://github.com/amazon-science/Spherical_Diffusion_Policy","path":null,"label":"amazon-science/Spherical_Diffusion_Policy","type":"external"},{"url":"https://github.com/amazon-science/TransitionFlowMatching","path":null,"label":"amazon-science/TransitionFlowMatching","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/improving-quality-and-robustness-in-llm-based-text-to-speech-systems","path":null,"label":"amazon.science/blog","type":"external"},{"url":"https://github.com/amazon-science/LARCQ","path":null,"label":"amazon-science/LARCQ","type":"external"},{"url":"https://github.com/amazon-science/omnimatch","path":null,"label":"amazon-science/omnimatch","type":"external"},{"url":"https://github.com/amazon-science/SWAN","path":null,"label":"amazon-science/SWAN","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://github.com/amazon-science/TurboFuzzLLM","path":null,"label":"amazon-science/TurboFuzzLLM","type":"external"},{"url":"https://github.com/amazon-science/JuLS","path":null,"label":"amazon-science/JuLS","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":18,"hiring":0,"talking":22,"releases":20},"data_radar_lanes":{"data":10,"evals":6,"safety":6,"product":3,"infrastructure":8},"data_radar_matches":23}},{"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}}]}