AI Lab Radar

Which AI labs ship — and when?

The AI frontier is shifting, and a large share of it now ships openly from China. The AI Lab Radar tracks which labs release new open-weight models: straight from Hugging Face, fetched again at least once a month.

Updated 30 Sept 2026Open-weight releases · refreshed at least monthly.
41model releases (90 days)
10active labs
61%of them from China

The model releases

New open-weight models, newest first. Filter by region. A click opens the model on Hugging Face.

All AI news in the ticker
  1. 20 Sept 2026Alibaba QwenQwen-Image-2.1-PE-T2Itext-to-image
  2. 20 Sept 2026Alibaba QwenQwen-Image-2.1-PE-I2I
  3. 16 Sept 2026Tencent HunyuanWeVisDoc-4B
  4. 14 Sept 2026Alibaba QwenQwen-Image-2.1text-to-image
  5. 10 Sept 2026DeepSeekDeepSeek-V4.1-Flashimage-text-to-text
  6. 8 Sept 2026NVIDIANV-Reason-CTimage-text-to-text
  7. 2 Sept 2026MicrosoftVibeVoice-ASR-Streaming-7Bautomatic-speech-recognition
  8. 1 Sept 2026NVIDIAKumo-Tabular
  9. 1 Sept 2026NVIDIANemotron-3-Diarizationvoice-activity-detection
  10. 31 Aug 2026DeepSeekDeepSeek-V4-Flash-Vision-Expimage-text-to-text
  11. 27 Aug 2026Tencent HunyuanHy4-previewtext-generation
  12. 27 Aug 2026Alibaba QwenQwen-Drive-1.0-4Bimage-text-to-text
  13. 25 Aug 2026Zhipu (GLM)GLM-5.3text-generation
  14. 25 Aug 2026Zhipu (GLM)GLM-5.3-Flashimage-text-to-text
  15. 24 Aug 2026Googletimesfm-3.0-pytorchtime-series-forecasting
  16. 24 Aug 2026Alibaba QwenQwen3.8-Flash-Nextimage-text-to-text
  17. 21 Aug 2026Tencent HunyuanAuK-Flashtext-to-speech
  18. 18 Aug 2026Tencent HunyuanAuKtext-to-speech
  19. 17 Aug 2026IBM Granitegranite-speech-5.0-470m-turboctc-ncautomatic-speech-recognition
  20. 17 Aug 2026Tencent HunyuanEVIE-Preview-4.5Bvisual-document-retrieval
  21. 14 Aug 2026Tencent HunyuanUI-Mate-27Bimage-text-to-text
  22. 13 Aug 2026DeepSeekDeepSeek-V4-Pro-0813text-generation
  23. 8 Aug 2026Alibaba QwenQwen3.8-2.4T-A95Btext-generation
  24. 7 Aug 2026IBM Granitegranite-4.2-30btext-generation
  25. 7 Aug 2026MiniMaxMiniMax-Music3text-to-audio
  26. 5 Aug 2026Alibaba QwenQwen3.8-27Bimage-text-to-text
  27. 4 Aug 2026IBM Granitegranite-speech-5.0-470m-turboctcautomatic-speech-recognition
  28. 31 Jul 2026DeepSeekDeepSeek-V4-Flash-0731text-generation
  29. 29 Jul 2026NVIDIANVIDIA-NemotronLabs-VoiceChat-11B
  30. 28 Jul 2026MiniMaxMiniMax-H3image-text-to-video
  31. 26 Jul 2026MicrosoftMage-ViTimage-feature-extraction
  32. 25 Jul 2026MicrosoftMage-VLimage-text-to-text
  33. 24 Jul 2026MicrosoftVibeVoice-ASR-BitNetautomatic-speech-recognition
  34. 17 Jul 2026MicrosoftFara1.5-27Bimage-text-to-text
  35. 16 Jul 2026Mistral AIShieldstral-1.0-3B
  36. 14 Jul 2026Tencent HunyuanHy-Embodied-RxBrain-1.0any-to-any
  37. 14 Jul 2026Tencent HunyuanHy-Embodied-VLM-1.0image-text-to-text
  38. 8 Jul 2026Tencent HunyuanR3-rerank-0.6btext-ranking
  39. 7 Jul 2026NVIDIACosmos3-Super-Image2Video-4Stepimage-to-video
  40. 6 Jul 2026NVIDIANemotron-Labs-Audex-30B-A3Btext-generation
  41. 2 Jul 2026Tencent HunyuanHy3text-generation

Size and quantization variants are collapsed to one release per model family; ♥ = traction (likes) on Hugging Face.

How we measure

Honest and verifiable — only the public Hugging Face API, no AI model involved.

We read the newest models of the major open-weight labs via the public Hugging Face API. We filter out quantizations, format ports, research artifacts and derivatives of other labs' base models, collapse size variants per model family, and show only releases with visible traction. Closed models such as GPT, Claude or Gemini don't appear here; find their announcements in the AI news ticker. We are not affiliated with any lab.

Tracked labs: DeepSeek, Alibaba Qwen, Moonshot (Kimi), Zhipu (GLM), Tencent Hunyuan, StepFun, MiniMax, Meta Llama, Google, Microsoft, NVIDIA, IBM Granite, Ai2, OpenAI, Mistral AI, Stability AI.

All information without guarantee. Data source: Hugging Face (public model API). "Open-weight" means the weights are freely available — not necessarily under an OSI licence. The lab selection is curated and not exhaustive. Model and company names belong to their owners. i6eal is not affiliated with any lab.

FAQ

What does the AI Lab Radar show?

Which AI labs release new open-weight models — with date, region and traction. A running chronicle of who is shipping what at the open AI frontier.

Why does so much come from China?

Chinese labs such as DeepSeek, Qwen, Kimi, GLM and Hunyuan release even their frontier models with weights. Western labs such as Google, Microsoft, NVIDIA, IBM, Meta and Mistral release open models too, but mostly keep their strongest models closed. Fully closed providers such as Anthropic don't appear here; their news is in the ticker.

What does "open-weight" mean?

The model weights are publicly downloadable — you can self-host, adapt and run the model without a third-party API. That isn't the same as "open source" (licences vary), but it's often decisive for data protection and independence.

How often does the radar update?

The list is rebuilt from the Hugging Face API at least once a month; the date is shown at the top. It always covers the 90 days before that date.

Open AI that runs at your place.

We bring open-weight models safely into your operations — selection, hosting, integration, training.