← Analysis archivereport.jsonevidence.jsonDossier →
InclusionAI (Ant Group)Neolabgenerated Oct 12, 2026 · 20h

InclusionAI (Ant Group) analysis

59 cited sources283 signals on file

Thesis

InclusionAI is Ant Group's open-weights 'AGI initiative' W5, running a three-family lineup — Ling (non-thinking), Ring (thinking), and Ming (multimodal) — released mostly under MIT/Apache-2.0 licenses W5W6E1E3E2. This pack shows a lab shipping foundation and vision-language models on a tight cadence E1E2E3E4E5E6E7E8E9E10E11E12E13 while, in parallel, building its own serving, sandboxing, and multi-agent infrastructure P5P8P9P10P16 and signaling EU regulatory readiness P13W3. The combined signal is a frontier lab investing in capability and the platform to operate it, with a staged GTM approach that sometimes ships evaluation-only artifacts before weights W2P4.

Signal desks

Hiring

  • No cited evidence in this pack (no open roles or job postings appear in the supplied evidence).

Forks

  • sglang (parent sgl-project/sglang), 3 stars — inference-runtime inspection E59.
  • vllm (parent vllm-project/vllm) — inference-runtime inspection E60.
  • Read: fork activity targets serving runtimes rather than models or agent frameworks, aligning with Awex's in-house sglang and Mooncake weight-transfer work P9P10.

Releases

  • HF model cards: Ling-3.0-tiny/flash/flash-VL/flash-Fin E1E3E5E6; LLaDA image + LLM family E7E8E9E10E12; Ming-Image-0.1-Design and Design-Layer E2E4; UI-Venus-2-9B E11; LemonDecider E13P1; AI-Transparency documentation E18P13.
  • singprobe sweep: small (~2–12M param) first-party finetune/probe releases across Qwen3.6/3.8 E25E38, Llama-3.2-1B E33, DeepSeek-V4-Flash E39, Step-3.7-Flash E45, gpt-oss-120b E46, MiniMax-M2.7 E47, GLM-5.3 E48, and Qwen3-0.6B E44.
  • Infra repo releases: AKernel sandbox runtime (rc6→rc9) P16P11P15P5; Avernet multi-agent platform v2026.09.17 P8; Awex weight-transfer/serving P9P10P20; Choruz builds + v0.2.3 library split P6P7P24; distill-fs P14P28; AEnvironment P3; humming P2.

Talking

  • Ming-Image framed as design-grade generation (UI/infographics/posters) with editable RGBA layers and agent skills Ling UI Design + Image-to-Editable-PPT W1W4.
  • PolyOCR-Venus framed as unified OCR; third parties flag that weights and training code are absent and only evaluation code ships W2.
  • EU AI Act Article 53(1)(d) training-content summaries for Ling/Ring/Ming published as documentation-only W3P13.
  • Third-party profiles: separate from Alibaba's Qwen team; trillion-scale FP8 training and AReaL async RL; MIT/Apache licensing as deliberate open-source strategy W5W6.

Shipping

Concrete artifacts in the pack:

  • Weights: Ling-3.0-flash (127.5B) E3, Ling-3.0-tiny (7.9B) E1, flash-VL E5, flash-Fin E6; LLaDA2.2-flash/mini and LLaDA image/UI models E7E8E9E10E12; Ming-Image pair (6.15B each) E2E4; UI-Venus-2-9B E11; LemonDecider (Qwen3.5-4B base, 255-way single-token readout, abstention) P1E13.
  • Repo releases: Avernet v2026.09.17 (717 PRs, 42 contributors) P8; Choruz v0.2.3 splits evaluation/community/learning libraries P24; Realtime-Venus Android beta APK P22; AKernel and Awex release cadence P5P9P20.
  • Not shipped: PolyOCR-Venus publishes a technical report and OCRBench evaluation code only — 'model weights and training code are not included' W2P4.

Research themes

  • Typed decision/distillation: LemonDecider distills Qwen3.5-4B into a no-autoregression, single-forward-pass decision head with an unknown/abstention candidate; tags include distillation and autojev P1E13.
  • Text-centric vision/OCR: PolyOCR-Venus 2B/9B OCR family on Qwen3.5 with OCRBench v2.1 / CC-OCR / OmniDocBench v1.6 results P4W2.
  • Visual design + layers: Ming-Image 6B text-to-image and layer-decomposition models with RGBA output P17P18P19.
  • Realtime multimodal: Realtime-Venus full-duplex audio/video with asynchronous delegation (Venus Team + Tsinghua; arXiv 2609.13814) P27P22.
  • Interpretability/probing: the singprobe finetune/probe series across a broad model zoo implies an internal probe/eval program E25E38E44E48.
  • Agent/sandbox systems: AKernel gVisor/runc sandboxes P5P16, Avernet multi-agent orchestration P8; Choruz v0.2.3 notes 'Jev integration is being researched and is not included', linking to the LemonDecider/autojev decision engine P24P1.

Hiring & scaling

No job-posting evidence in this pack, so hiring teams, locations, or functions cannot be derived. Scaling is only indirectly visible: Avernet merged 717 PRs from 42 contributors P8 and Choruz ships near-daily verified builds on a Linux x86_64 / Node 24 / PostgreSQL baseline P6P7P12. Treat any org-size inference as unproven beyond these signals.

Category implications

  • Strategy: open-weights positioning as an 'AGI initiative' distinct from Alibaba's Qwen, with MIT/Apache licensing framed as a competitive lever W5W6.
  • Infrastructure: AKernel (sandbox runtimes, ARM64, gVisor/runc) P5P16, Awex (Mooncake, sglang, Qwen4Exp converters) P9P10P20, and Avernet/Choruz indicate a build-out of proprietary agent-serving and sandbox substrate.
  • Product: agent-skill packaging (Ling UI Design, Image-to-Editable-PPT) W1W4 and Realtime-Venus async delegation P27 point toward agent/creator tooling rather than bare model cards.
  • Research: singprobe probes E25E44E48 and LemonDecider P1 imply work on decision heads/interpretability that could feed evaluation and reliability tooling; Choruz's in-progress Jev integration P24 reinforces this.
  • Hiring: no evidence — do not infer team composition beyond the contributor signals P8.
  • GTM: EU AI Act training-content summaries P13W3 signal a regulatory-readiness posture for the EU; PolyOCR-Venus shipping evaluation-only code W2 is a staged release that withholds weights/API.

Traction highlights

  • Ling-3.0-tiny: 26,419 downloads / 448 likes E1; Ling-3.0-flash: 17,049 / 406 E3; Ling-3.0-flash-VL: 8,801 / 102 E5.
  • Ming-Image-0.1-Design: 415 likes E2; Design-Layer: 99 likes E4; Ming-Image repo logged 192 stars at repo_new E36.
  • LLaDA2.2-flash: 1,744 downloads / 89 likes E8; LLaDA-Image: 770 / 81 E7.
  • Realtime-Venus repo: 141 stars at repo_new E40; PolyOCR-Venus: 50 stars E14.
  • The singprobe long tail shows minimal adoption (roughly 1–5 likes, ~130–580 downloads) E25E33E44E54, consistent with research probes rather than consumer releases.