Introducing Mistral Large 4: Le chonk
Mistral’s new preview is a 1T-parameter model running through its API, with open weights promised by month’s end.
Simon Willison says Mistral Large 4 uses 49B active parameters and was trained on Mistral’s 3,800 Grace Blackwell GPU cluster. Its API exposes only two reasoning settings, “none” and “high,” and the higher setting produced the better pelican image in his test. On Artificial Analysis it scores 38, far above Mistral Large 3’s 9, but still behind DeepSeek 4.1 Flash. Simon Willison's note
Simon Willison says Mistral Large 4 uses 49B active parameters and was trained on Mistral’s 3,800 Grace Blackwell GPU cluster. Its API exposes only two reasoning settings, “none” and “high,” and the higher setting produced the better pelican image in his test. On Artificial Analysis it scores 38, far above Mistral Large 3’s 9, but still behind DeepSeek 4.1 Flash. Simon Willison's note
score 8