August’s roundup was about flagship weights that came with bespoke licences. September went mostly the other way. The largest releases, from Xiaomi, Tencent, DeepSeek, and Shanghai AI Lab, came under MIT or Apache 2.0, and NVIDIA used its permissive OpenMDW licence. The traps moved to smaller, specialist models. Alibaba’s Qwen-Image-2.1 and Cohere’s North Small Translate are both non-commercial, even though both appear in the same “open” listings as the MIT flagships.

What happened

Xiaomi MiMo-V2.6 (21 September, MIT)

Xiaomi published three checkpoints on Hugging Face:

  • MiMo-V2.6-Pro-RL: a 1.02-trillion-parameter mixture-of-experts model with 42B active. It has 384 routed experts, with 8 active per token.
  • MiMo-V2.6-Flash-RL: 310B total and 15B active.
  • MiMo-V2.6-Distill-Qwen-9B: a supervised fine-tune of Qwen3.5-9B on MiMo-generated data. Xiaomi releases it as a starting point for agentic-RL research.

Pro and Flash are omnimodal (text, image, video, and audio input) with a 1M-token context. Xiaomi describes a single mixed RL run across coding, agents, vision, and cybersecurity. On its own evaluations, Pro scores 71.9 on DeepSWE v1.1 against Claude Opus 5’s 74.0. The self-reported CyberGym scores are 94.0 for Pro and 95.1 for Flash. Taken at face value, that is among the highest cyber scores of any downloadable model.

Tencent Hy4-preview (27 August, Apache 2.0)

Tencent’s Hy team posted Hy4-preview the day before our August roundup was published, and it was not covered there. It is a 770B-parameter MoE with 49B active: 256 routed experts plus 1 shared, with the top 8 active per token. It has a 1M-token context and a built-in multi-token-prediction layer for speculative decoding. Tencent calls it a preview. Expect a final checkpoint, which may differ.

DeepSeek-V4.1-Flash (10 September, MIT)

DeepSeek-V4.1-Flash is a 552B multimodal MoE with 8B active in prefill and 16B in decode. Its KV cache is about a quarter the size of V4-Flash’s, and it is licensed under MIT. See the full write-up . NVIDIA followed with an MIT-licensed NVFP4 quantization on 16 September.

Qwen-Image-2.1 (14 September, Qwen Research Licence)

Qwen-Image-2.1 combines text-to-image generation and image editing in one model, with a 7B DiT visual generator. It can output transparent (RGBA) images and accepts up to 10 reference images for editing. Hugging Face’s metadata shows the licence only as “other”. The LICENSE file is the Qwen RESEARCH LICENSE AGREEMENT. It grants rights “FOR NON-COMMERCIAL PURPOSES ONLY”, defines non-commercial as “research or evaluation purposes only”, and says commercial use requires a separate licence from Alibaba. That is a departure from the Apache 2.0 licence on Alibaba’s recent small LLMs, such as Qwen3.8-27B.

Cohere North Small Translate (9 September, CC BY-NC 4.0)

North Small Translate (north-small-translate-1-0) is a translation-specific MoE:

  • 218B total and 25B active parameters
  • 16K context
  • 50+ languages and locale variants
  • Free tier on Cohere’s Chat V2 API
  • Weights in W4A16, FP8, and BF16 under Creative Commons Attribution-NonCommercial 4.0

The Hugging Face repo is gated. Cohere suggests two H100s or one B200 for W4A16. Cohere’s small tiny-aya “thinker” models (early September) are also CC BY-NC.

NVIDIA Nemotron-3-Labs-Ultra-Math (3 September, OpenMDW-1.1)

NVIDIA published Nemotron-3-Labs-Ultra-Math-RL and -SFT. They are math-reasoning and proof-checking specialists built on Nemotron-3-Ultra, a hybrid Mamba2-Transformer LatentMoE with 550B total and 55B active parameters. The model card says the RL checkpoint was “deployed as part of an ensemble system” that reached a gold-medal-level score at IMO 2026. That gold was scored by the ensemble, not by this checkpoint on its own. The licence is OpenMDW-1.1, and the card states the model is “ready for commercial and non-commercial use”. The recipe is in arXiv 2609.10712. See NVIDIA Nemotron 3 for the base family.

Shanghai AI Lab Intern-S2-397B (13 September, Apache 2.0)

Intern-S2-397B is a multimodal model for scientific reasoning and long-horizon agents. It is pre-trained directly on raw pages of scientific literature and trained with RL across more than 20 scientific domains. The same lab’s Atria-Dawn-Preview (11 September, MIT) is an agentic model. It has FP8 and Ascend W8A8 variants.

GLM-5.3-Flash: check which GLM you mean

GLM-5.3-Flash (320B total, 18B active, natively multimodal) is MIT-licensed. The full GLM-5.3 from August is not. It uses Z.ai’s bespoke GLM-5.3 License, with a security-review condition for large Model-as-a-Service operators. Same family name, different terms.

Smaller releases

  • inclusionAI Ling-3.0-flash-VL (4 September, MIT): 124B total and 5.5B active, image and video input, 256K context.
  • OpenBMB MiniCPM5-2B (6 September, Apache 2.0).
  • Liquid AI LFM2.5-VL-3B-DSpark (18 September) uses Liquid’s own LFM 1.0 licence. It limits commercial use by organisations above a $10M annual revenue threshold.
  • Qwen-Drive-1.0-4B (27 August, Apache 2.0): a driving-domain model.

Licence summary

ModelSize (total / active)LicenceCommercial use
MiMo-V2.6-Pro-RL / Flash-RL1.02T / 42B; 310B / 15BMITYes
Tencent Hy4-preview770B / 49BApache 2.0Yes
DeepSeek-V4.1-Flash552B / 8B–16BMITYes
Intern-S2-397B397BApache 2.0Yes
Nemotron-3-Labs-Ultra-Math550B / 55BOpenMDW-1.1Yes
GLM-5.3-Flash320B / 18BMITYes
GLM-5.3 (August)753BGLM-5.3 License (bespoke)Yes, with conditions
Ling-3.0-flash-VL124B / 5.5BMITYes
Qwen-Image-2.17B generatorQwen Research LicenceNo, separate licence needed
North Small Translate218B / 25BCC BY-NC 4.0No
LFM2.5-VL-3B-DSpark3BLFM 1.0Restricted above $10M revenue

Why it matters for builders

Hugging Face metadata is not a licence review. Qwen-Image-2.1 shows up as “other”, and you only find “research or evaluation purposes only” in the LICENSE file. North Small Translate is available on a free API tier, which makes it easy to assume the weights are free to use commercially. They are not. Make “open the LICENSE file” a required step in model intake. See software licensing and vendor lock-in .

The permissive frontier is now very large and very cyber-capable. MIT and Apache 2.0 now cover 500B–1T-parameter models with self-reported CyberGym scores near or above 90 (MiMo-V2.6, DeepSeek-V4.1-Flash). Security teams should assume attackers can download these capabilities and use them without restriction. This is the downloadable counterpart to the gated cyber tiers closed labs launched this month.

Sizes still decide who can run what. Every flagship here needs multi-GPU serving. The practical self-hosting tier is still around 30B total or less, or low-active MoEs like Ling-3.0-flash-VL. Quantization (NVIDIA’s NVFP4 builds, Cohere’s W4A16) is how the larger models become deployable. See small vs large language models .

Every benchmark above is self-reported. Labs used different harnesses and effort settings. Use the numbers as directional and run your own evaluations .

Sources

  1. Xiaomi MiMo, “MiMo-V2.6-Pro-RL” model card and technical report (21 September 2026): https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL
  2. Xiaomi MiMo, “MiMo-V2.6-Distill-Qwen-9B”: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
  3. Tencent Hy, “Hy4-preview” (27 August 2026): https://huggingface.co/tencent/Hy4-preview
  4. DeepSeek, “DeepSeek-V4.1-Flash” (10 September 2026): https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
  5. Qwen, “Qwen-Image-2.1” model card and LICENSE (14 September 2026): https://huggingface.co/Qwen/Qwen-Image-2.1
  6. Qwen, “Qwen-Image-2.1” blog: https://qwen.ai/blog?id=qwen-image-2.1
  7. Cohere, Changelog, “Announcing Cohere’s North Small Translate” (9 September 2026): https://docs.cohere.com/changelog
  8. NVIDIA, “Nemotron-3-Labs-Ultra-Math-RL” (3 September 2026): https://huggingface.co/nvidia/Nemotron-3-Labs-Ultra-Math-RL
  9. NVIDIA, “An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics”, arXiv 2609.10712: https://arxiv.org/abs/2609.10712
  10. Shanghai AI Lab, “Intern-S2-397B” (13 September 2026): https://huggingface.co/internlm/Intern-S2-397B
  11. Z.ai, “GLM-5.3-Flash”: https://huggingface.co/zai-org/GLM-5.3-Flash
  12. inclusionAI, “Ling-3.0-flash-VL” (4 September 2026): https://huggingface.co/inclusionAI/Ling-3.0-flash-VL
  13. Liquid AI, “LFM2.5-VL-3B-DSpark” and LICENSE (18 September 2026): https://huggingface.co/LiquidAI/LFM2.5-VL-3B-DSpark
  14. NVIDIA, “DeepSeek-V4.1-Flash-NVFP4” (16 September 2026): https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-NVFP4

Further reading