Signum
Feed
Strong signal29 Oct 2026high confidence

Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support

Mistral Large 4 ("Le Chonk") weights were published on Hugging Face (model id mistralai/Mistral-Large-4-1T-A52B) in FP8 and NVFP4 official variants, plus GGUF quants from Unsloth and the llama.cpp team. It is a multimodal (text+image in, text out) MoE with about 1T total and 49B active parameters, a 1M-token context window, and a Pixtral 1.6B vision encoder. It uses MLA plus a DeepSeek Sparse Attention indexer. The licence is the Mistral Ecosystem Open-Weight License, an MIT derivative restricting model-as-a-service offerings. Day-0 support covers transformers, vLLM and llama.cpp, along with TRL fine-tuning, Inference Endpoints, Inference Providers and Dell Enterprise Hub. Reported results: 82% on the Artificial Analysis Cyber Index reproduce-and-patch test, 93% on Cybench, 61.7% on DeepSWE v1.1, and fifth-best open-weights model on the AA Intelligence Index as of Oct 29, 2026. Most benchmark numbers are Mistral-reported or quoted by Hugging Face, and the post does not say which were independently verified.

CapabilityAccessInfrastructure

Entities: Mistral AI, Mistral Large 4, Hugging Face, Artificial Analysis, Unsloth, llama.cpp

80Strong signal
1 source
1 primary
Was this useful?
01

What happened

Mistral published the weights for Mistral Large 4 on Hugging Face (mistralai/Mistral-Large-4-1T-A52B) on 29 October 2026, in official FP8 and NVFP4 variants. Community GGUF quantisations from Unsloth and the llama.cpp team are also available. It is a multimodal mixture-of-experts model: text and images in, text out, about 1T total parameters with 49B active per token, and a 1M-token context window. Support for transformers, vLLM and llama.cpp is available from day zero. The licence is an MIT derivative that bars offering the model as a hosted service.

02

Why it matters

Teams that want to self-host a frontier-scale model now have another option, with ready tooling for fine-tuning and deployment, including Dell Enterprise Hub. The strongest practical case is defensive security work, where hosted closed models often refuse requests. Companies planning to resell access to the model as an API cannot do so under this licence. Running a 1T-parameter model also needs serious hardware even with FP8 or 4-bit weights, so many smaller teams will not be able to use it directly.

03

What is noise

The headline results (82% on the Cyber Index reproduce-and-patch test, 93% on Cybench, 61.7% on DeepSWE v1.1) are mostly reported by Mistral or relayed by Hugging Face, and the post does not say which were independently verified. Phrases such as a "huge boost for the community" are promotion from a distribution partner, not evidence. "Open-weight" is also not "open source": the usage restriction matters commercially. Fifth-best open-weights model on one index is solid but not a leap.

04

Watch next

  1. 01Independent replication of the Cybench and DeepSWE numbers, and the model's position on third-party leaderboards such as Artificial Analysis and LMArena, over the next two to four weeks
  2. 02Real-world throughput, memory needs and cost per token reported by vLLM and llama.cpp users, and which hosting providers offer it despite the licence restriction
  3. 03Clarification or legal reading of the Mistral Ecosystem Open-Weight License terms on model-as-a-service use, plus any misuse reports linked to its cyber capabilities

Coverage

1 story

Newsletter

Get the no-hype AI news straight to your inbox

Weekly. No spam, unsubscribe any time. Privacy

More capability signals

Full feed →