Mistral Large 4 goes open-weights: 1T parameters, trained in Europe

Mistral published the weights of Mistral Large 4 on 29 October 2026: a multimodal mixture-of-experts model with a 1M-token context, 160+ languages and a license with a hosting condition.

  • New release
  • 2 min read
  • Published
  • Updated

Mistral AI published the weights of Mistral Large 4, nicknamed Le Chonk, on 29 October 2026. The company had previewed the model on its API on 6 October and trained it for three more weeks with reinforcement learning before releasing the weights. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers.

What is new

  • Architecture. A natively multimodal mixture-of-experts model with about 1 trillion parameters in total and 49 billion active per token (52 billion if you count the embeddings and the output layer). It has 128 routed experts with 4 active for each token, plus 1 shared expert.
  • Inputs. It accepts text and images, writes text, and has a 1-million-token context window. A Pixtral-style vision encoder handles images of different sizes.
  • Languages. Mistral says it speaks over 160 languages, including every official language of the European Union.
  • Reasoning. Three reasoning-effort settings, low, medium and high, trade speed for depth. The checkpoint includes a built-in draft head for faster speculative decoding.
  • Focus. Mistral emphasizes coding and defensive cybersecurity. Its own report lists 19.7 percent on Cognition’s private FrontierCode 2.0 and 41 percent on Terminal-Bench 4.0, which are vendor-reported figures.

Running it

The official checkpoint is about 1.06 TB. An NVFP4 version of about 610 GB targets NVIDIA Blackwell GPUs and works with vLLM and SGLang. Either way, this is a multi-GPU server model, not something for a desktop. If you are choosing between large open models from outside China, see our comparison with Muse Spark Open.

License

The Mistral Ecosystem Open-Weight License is a modified MIT license. Anyone may use the model and sell products built on it, but a business that offers hosted access to the model to third parties needs a separate commercial license from Mistral. There is no revenue threshold, so a small hosting startup is covered as much as a large one. Building the model into your own product, or relaying requests to a model that someone else hosts, does not count as hosting.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Mistral Large 4 Mistral AI 1.05T 52B active 1M Mistral Open-Weight (Conditional) 671 GB GPU server