Mistral has released the full weights of Mistral Large 4, the model it nicknames 'Le Chonk', on Hugging Face. The release follows an October preview and three extra weeks of reinforcement learning, and it makes Large 4 the largest open-weight model any US or European lab has published.
The model is a mixture-of-experts design with about 1.05 trillion parameters in total and roughly 52 billion active per token. It handles text and image input, offers a one-million-token context window, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral's own European data centres — a point the company leads with when it sells the model as a foundation for 'sovereign AI'. Mistral also cut its hosted API price roughly in half on release day, to about $0.70 per million input tokens.
Open weights, with fine print
Large 4 ships under a new Mistral Ecosystem Open-Weight License, a modified MIT licence. Companies can download, modify, fine-tune and self-host the model for their own products and internal tools. But a business that sells hosted access to the model — a 'Model as a Service' offering — must sign a separate commercial licence first.
That single clause is aimed squarely at cloud and inference providers, and it marks a shift for open models: the weights are public, the training data is not, and the right to resell the model as a service stays with Mistral. For everyone else, the practical limit is hardware — at this size, Large 4 is a data-centre model, not something that runs on a laptop.
The bigger picture
Mistral's position in the AI race is distinctive: it is the leading European frontier lab, and it sells that identity as hard as it sells the technology. Training Large 4 from scratch on thousands of GPUs in its own European data centres is a sovereignty argument — proof that a frontier-class model can be built outside the American and Chinese ecosystems, on infrastructure Europe controls.
The licence, meanwhile, is a business model in legal form. Free for anyone to run themselves; paid for anyone who wants to sell it as a service. That split lets Mistral claim the open ecosystem's goodwill and developer mindshare while reserving the most lucrative use — hosted inference at scale — for its own commercial agreements. It is the same tension the whole open-model movement is negotiating, written into a single clause.
What happens next
The practical test is adoption: whether the open ecosystem builds on Large 4 despite the service clause, and whether cloud providers sign the commercial licences or steer customers to models they can host freely. Either way, the era of simply calling a release 'open' is ending — Large 4's fine print is what the word now has to survive.
Common questions
How big is Mistral Large 4?
Large 4 is a mixture-of-experts model with about 1.05 trillion parameters in total and roughly 52 billion active per token. It handles text and image input, offers a one-million-token context window, and was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centres.
What is the catch in Large 4's open licence?
The Mistral Ecosystem Open-Weight License allows free internal use — companies can download, modify, fine-tune and self-host the model. But selling hosted access to it as a 'Model as a Service' offering requires a separate commercial licence from Mistral.




