Mistral released the weights of Mistral Large 4 with a trillion parameters

On October 30, 2026, the French company Mistral released the weights of the Mistral Large 4 model. Size — a trillion parameters (a parameter is a numerical coefficient that the model adjusts during training; the more of them, the heavier the model). The weights are available on Hugging Face and via the Mistral API. Companies and government agencies can download them and run them locally.

The model has the nickname "Le Chonk". It was introduced earlier this same month, and now the weights have been opened. This is a continuation of Mistral's line of open-weight models (open-weight — the weights are published, but not necessarily the training data and code). The organization keeps the data and infrastructure itself rather than going to an external provider.

The difference from the previous step is that previously access to the model went through an API, that is, through someone else's servers. Now the weights can be taken for oneself. Mistral positions this as an alternative to closed developers like OpenAI and Anthropic: for them, access and operation go through an external supplier. For cybersecurity this is an argument: where sensitive data is processed, it is more convenient not to send it outside.

CEO Arthur Mensch describes the goal more broadly — to make AI "a tool for empowerment, not domination". Mistral claims that the model delivers a state-of-the-art result with less compute and capital than closed developers. There are no specific comparison figures in the announcement.

In practice, this is a variant for those who want to keep the model on their own hardware: code analysis, vulnerability discovery, incident investigation. Organizations are increasingly looking at AI specifically for these tasks. But a trillion parameters is a heavy model, and the hardware requirements are not stated in the announcement. Downloading the weights and running them are different things.

What remains unknown: what data the model was trained on, under what license the weights are released, and what resources are needed to run it. Without this, the claim of a "state-of-the-art result with fewer resources" cannot be evaluated.