Mistral Large 4 'Le Chonk' open weights have landed, under a new modified-MIT licence

When Mistral announced Large 4 in early October it put the model behind an API preview and said the open weights would follow “by the end of the month”. They have. On 29 October Mistral published the weights for Mistral Large 4, the trillion-parameter multimodal model it nicknames “Le Chonk”, and the most interesting part of the release is the licence it arrived under.

What landed

The canonical repository is mistralai/Mistral-Large-4-1T-A52B, with an NVFP4 variant for Blackwell GPUs alongside it. It is a hybrid instruct-and-reasoning mixture-of-experts of around one trillion total parameters: 72 decoder layers (the first dense, the rest MoE), a router that picks four of 128 experts plus one always-on shared expert, and about 49 billion active parameters per token, which is where the “A52B” in the repository name comes from once embeddings are counted. It takes text and images in, using a 1.6-billion-parameter Pixtral vision encoder, and produces text, with a context window up to one million tokens, though the deployment examples commonly cap at 131k.

Mistral ships FP8 and NVFP4 checkpoints, and the llama.cpp team and Unsloth have published community GGUF builds down to 1- and 2-bit. For serving, the announcement points at vLLM, SGLang, an NVIDIA NIM container, and TensorRT-LLM.

The licence is the news

Large 3 was released under Apache 2.0. Several of Mistral’s bigger models before it used the non-commercial Mistral Research License. Large 4 takes a third path: a new Mistral Ecosystem Open-Weight License, which Mistral describes as “a modified version of the MIT license”. We read the licence file rather than the summaries, and the shape is genuinely permissive:

• It grants broad rights to use, copy, modify, fine-tune, distribute, sublicense and sell, and to run, deploy and further train the model.
• The one substantive condition is on Model-as-a-Service. If you or an affiliate operate a Model-as-a-Service business, meaning you give third parties inference or post-training access with meaningful control over inputs, parameters or training data, you must obtain a separate commercial licence from Mistral before any commercial use.
• That condition explicitly does not apply to internal use, or to embedding the model inside an end-user product where its capabilities sit behind specific features. There is no revenue threshold and no territory clause, and the only other requirement is the MIT-style one of keeping the copyright notice.

In our catalogue terms that is permitted-conditional, but at the permissive end of it. For almost everyone except a company whose business is hosting the model for others, this behaves like an open commercial licence. It is a more open stance than the conditional licence on MiniMax M3.1 Flash last week, which asked for attribution and a notice even for ordinary commercial use, and a clear step back toward the openness of Large 3 and Mistral’s Apache-licensed releases. The broader theme we keep returning to, that open weights and “freely usable” are not the same thing, still holds, which is exactly why reading each licence rather than assuming matters.

What it takes to run

This is a large-memory model, not a consumer-card one, and the memory you need is set by the full trillion parameters, not the 49 billion active. Using the community GGUF sizes as a guide, a 1-bit build is around 233GB and a 2-bit build around 300 to 380GB, which is the single-machine route: it fits a 512GB Apple Silicon Ultra. A sensible-quality 4-bit build is around 605GB and needs roughly 768GB or more, or multi-GPU distributed inference, and the near-full-precision 8-bit checkpoint runs to about 1.12TB, which is multi-node server territory. All of these figures are approximate and rise with context and concurrency.

The headline is that a prosumer machine, a maxed-out Mac Studio, can run a trillion-parameter model at all, albeit at a 2-bit quantisation that is a real compromise for a frontier model. If you want it at a quality you would actually deploy, plan for a multi-GPU server. Full requirements, the quantisation table and the licence conditions are on the Mistral Large 4 catalogue page. As with any fresh release, the capability figures circulating now are Mistral’s own, and we will wait for independent evaluation before treating them as settled.