InclusionAI (Ant Group)

Thesis

InclusionAI (Ant Group's AI research division W3) is executing a full-stack, open-weight agent strategy rather than a single-model program. It ships frontier-scale Mixture-of-Experts and diffusion models under permissive licenses — Ling 3.0 Flash (124B total / 5.1B active) E4W2, Ling 3.0 Tiny (7.9B / ~1.3B active) E3W3, LLaDA2.2-mini under Apache-2.0 P5E13, and the 6B LLaDA-Image/Turbo pair under Apache-2.0 P8E10W1 — while simultaneously building the surrounding agent stack: sandboxes (AKernel) P16P24, environments (AEnvironment, AWorld) P19P20, training/serving (AReno, Awex) P6P27, intrinsic safety guardrails (SingProbe) P13P14, and a human-agent collaboration surface (Choruz) P10P3. The public framing is explicitly "inclusive AGI" — open weights and low-cost, useful access E58W5 — with finance as the clearest vertical anchor via Ling-3.0-flash-Fin P7E6.

Signal desks

• Hiring: No cited evidence in this pack. The supplied events are model releases, repo events, forks, and blog posts (e.g., E1, E4, E24, E56, E58); none encode job roles, teams, locations, or job descriptions.
• Forks: Two inference-serving forks — sgl-project/sglang E56 and vllm-project/vllm E57. Both align with published serving requirements: SingProbe is integrated via SGLang and vLLM branches P14P15, the DSpark speculator is served with SGLang P23, and Awex/asystem-awex implement SGLang sharding and weight exchange P18P27.
• Releases: A dense multi-family stream across Aug–Sep 2026: Ling 3.0 Flash/Tiny plus base variants E3E4E23E28E30E31E32E33, LLaDA2.2-mini E13, LLaDA-Image/Turbo E10E11, UI-Venus-2-9B E12, Ling-3.0-flash-Fin E6, SingProbe guardrails E14E16, DSpark E17, and infra releases (Choruz, AKernel, AReno, Awex, Avernet, cuLA) E19E20E21E22E25E29E34E35E37E46E47E48E50.
• Talking: Three inclusion-ai.org posts frame the mission ("inclusive AGI," agentic-ecosystem observations, and a Ming-Omni-TTS note) E58E59E60, while external coverage centers on the open, low-cost Ling models and the LLaDA-Image launch W1W2W3W5W6.

Shipping

The lab ships both weights and systems. On models: Ling 3.0 Flash — a 124B MoE activating 5.1B per token, positioned for production-scale agentic workloads — and Ling 3.0 Tiny (7.9B, ~1.3B active, native 256K context, function calling, prompt caching) shipped in Aug 2026 E4W2E3W3. The diffusion line advanced with LLaDA2.2-mini, which adds Levenshtein Editing (DELETE/INSERT control tokens) for long-context tool calling and multi-turn correction P5E13. LLaDA-Image (6B) and a 4-step Turbo checkpoint shipped with fully open training recipes, BF16/FP8 variants, and text-to-image/editing/bilingual-text tasks P8W1E10E11. Domain and capability releases include Ling-3.0-flash-Fin (finance) P7E6, UI-Venus-2-9B (GUI agent) P17E12, SingProbe guardrails P14P15E14E16, ArmorOCR E15, and the DSpark speculative-decoding draft P23E17.

On tooling: Choruz went from first public release v0.1.0 to v0.1.2 plus nightly prerelease builds within the same week, adding semantic component-activity capture with authentication fields excluded P12P3P1P2. AKernel shipped v0.1.5 (Dockerfile-based sandbox launches, runc support, observability) and v0.1.6rc1 (Firecracker microVM support, same-node sandbox recovery) P24P16. AReno v0.0.8 added an Apple Silicon MLX backend and expanded multimodal, LoRA, and RL support P6. Awex v0.8.1 added Qwen3/Qwen3-VLM/Qwen3.5/Qwen3.6 weight conversion P27.

Research themes

1. Diffusion/editing-based language models. LLaDA2.2 introduces DELETE/INSERT control tokens (Levenshtein Editing) to improve long-context tool calling, multi-turn interaction, and error correction, with a marked BFCL v4 function-calling jump (47.68 vs 28.44 for 2.1-mini) P5. 2. Efficient MoE + speculative decoding. Ling 3.0 Flash activates ~1/64 of parameters per token W2; Ling 3.0 Tiny activates ~1.3B of 7.9B W3; DSpark adds a 1.36B draft with a confidence head, reaching a 5.29 macro-mean acceptance length across workloads P23. 3. Intrinsic streaming guardrails. SingProbe reuses frozen base-model hidden states to score query intent, response safety, and hallucination at every token with <0.5% decode overhead — 5.18M probe params for Flash, 3.22M for Tiny P13P14P15. 4. Scalable GUI agents. UI-Venus-2 spans 170+ apps, 50k+ websites, and desktop OS, with RL reward signals from visual keypoints and multi-model voting to resist reward hacking P17. 5. Open, reproducible training recipes. LLaDA-Image emphasizes fully open training recipes (220M samples, 98M real images per external coverage) P8W1, and AReno exposes SFT/DPO/GRPO/GSPO/PPO plus multimodal training across CUDA and MLX P6. 6. Financial domain specialization. Ling-3.0-flash-Fin is continued-trained on financial data with FinFIRST open-sourced for evaluation, targeting end-to-end financial research and valuation/spreadsheet workflows P7. 7. Train–infer weight exchange. asystem-awex implements NCCL weight exchange, SGLang/Megatron sharding, and weight-diff validation, pointing to tight train/inference co-optimization P18.

Hiring & scaling

No cited evidence in this pack. The evidence contains no open roles, team names, locations, or job descriptions — the feed is model cards, release notes, repo metadata, forks, and blog posts E1E4E24E56E58. Scaling direction can only be inferred indirectly from the breadth of shipped systems (sandboxes, environments, training/serving, guardrails) P6P16P19P20, but no direct hiring signal confirms team buildout.

Category implications

• Strategy: The "inclusive AGI" positioning and open-weight cadence reframe competition away from keynotes toward cheap, accessible, useful models E58W5. Finance is a deliberate vertical: Ling-3.0-flash-Fin extends the flash base via continued training with leading financial institutions, tying the lab to Ant Group's fintech context P7W3.
• Infrastructure: AKernel (Firecracker, runsc/runc, Dockerfile sandboxes, failover) P16P24, AEnvironment (SWE-bench, sandbox abstraction) P19P22, and AWorld (sandbox, swarm, memory, event-driven) P20 signal a serious bet on running untrusted agent code at production scale. asystem-awex and Awex indicate investment in MoE weight exchange/sharding and multi-framework (SGLang/Megatron) conversion P18P27.
• Product: Choruz is a local-first collaboration space that treats external agent CLIs (Claude Code, Codex, Pi, Grok, OpenCode) as first-class workers — a product-layer bet on multi-agent orchestration rather than a single proprietary agent P10P3. PanelWise extends this to self-hosted, evidence-grounded multi-model deep research E53.
• Research: Diffusion LLMs and intrinsic probes are differentiated, non-mainstream research bets with inspectable artifacts (technical reports, training code) P5P13P14.
• Hiring: No direct evidence; the systems-plus-agent breadth implies sustained ML-systems and infra hiring needs, but this is not confirmed by cited evidence P6P16.
• GTM: Distribution is multi-channel — Hugging Face, ModelScope, OpenRouter free tiers, and SGLang cookbooks — with permissive licenses (MIT/Apache-2.0) lowering adoption friction P7P8W2W3W5P23.

Traction highlights

• Model pulls: Ling-3.0-tiny leads at 26,539 downloads / 429 likes E3; Ling-3.0-flash at 18,100 / 403 E4; Ling-2.6-flash at 2,415 / 506 E2; Ling-3.0-flash-dspark at 2,867 downloads E17; UI-Venus-2-9B at 3,224 downloads E12. Newer releases show early low counts (LLaDA2.2-mini 46/12; LLaDA-Image 343; Turbo 571) E13E10E11.
• External attention: a Medium explainer on Ling 3.0 Flash W2, an independent review scoring Ling 3.0 Tiny 25 on Artificial Analysis' Intelligence Index (6th of 56 in its size class) W3, practitioner posts framing the lab's open/low-cost drumbeat W5W6, and an AI/TLDR writeup of LLaDA-Image W1.
• Repo stars: repo-creation events recorded Choruz at 410 stars E24 and LLaDA-Image at 191 E27; later metadata crawls list 61 P10 and 39 P11, respectively. PanelWise (138) and ConceptEdit (37) show early community interest E53E39.