Mistral releases Mistral Large 4

Mistral AI put Mistral Large 4 into public preview, a mixture-of-experts model with 1 trillion total parameters and 52 billion active per token. That is a step up from Mistral Large 3, released ten months earlier with 675 billion total parameters, and the company’s largest flagship to date. Mistral said the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, framing the choice as a sovereignty argument — a frontier-scale model built and run “independently of other digital service providers and under European law.”

The company leaned heavily on cases where it argued rivals’ built-in safety restrictions get in the way of legitimate professional work, particularly cybersecurity. Mistral said the model ranked among the top five globally on the AA Cyber Index, a third-party cybersecurity benchmark, and scored 82% on a vulnerability-reproduction test, contrasting this with what it called the “provider-level refusals” of other labs’ models when used by security researchers. It also reported 61.7% on the DeepSWE v1.1 coding benchmark, which it said beat other open-weight models, and claimed the model narrowly outperformed OpenAI’s GPT-6 Astra on a visual-grounding benchmark called Dense 200. These figures came from Mistral’s own testing rather than independent evaluation.

The model is available now through the Mistral Studio API, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral said open weights would follow by the end of October 2026.

Mistral was one of two Western developers to announce an open-weight model of more than 500 billion parameters that week; the other was Reflection AI with Beam.