Europe’s best-known AI lab has put its biggest model on the table for anyone to download. Mistral has released the open weights of Mistral Large 4, a trillion-parameter system nicknamed Le Chonk, after three extra weeks of training following its October preview.
What Has Actually Been Released
The weights were published on 29 October and are free to download, run and adapt under a modified MIT licence, with one catch: companies that want to run the model as a paid service to others need a separate commercial licence. Only a fraction of the model’s parameters, roughly 50 billion, are active for each token, a sparse design intended to keep running costs down, and the system is multimodal and fluent in more than 160 languages. A plain-English explainer of the release walks through what the download means for ordinary AI users as well as developers.
The Performance Claims, With a Caveat
Mistral says the finished model now matches the leading open model on a well-known coding-agent benchmark while using considerably fewer tokens per task, and the company is positioning it as the strongest open-weight option for coding and cybersecurity work. Those are vendor numbers, produced on the company’s chosen tests, and independent replication is still thin. Even so, the release landed at the top of developer discussions, because open models that genuinely rival the best closed systems remain rare.
Why Open Weights Matter
Releasing weights changes who controls the technology. A hospital, bank or government agency can run the model on its own servers, inspect how it behaves and keep sensitive data in-house, the same instinct that has pushed open models into everything from research labs to AI search tools answering everyday questions. Mistral also trained the system in its own European data centres on thousands of Nvidia Grace Blackwell chips, a sovereignty angle the Paris-based company is keen to emphasise.
Whether Le Chonk keeps its coding crown once independent testers finish with it is an open question. But a trillion-parameter European model that anyone can download is, by itself, a significant moment for open AI.
