Mistral Large 4 preview: a trillion-parameter model with open weights promised

Mistral AI has released a preview of Mistral Large 4, a trillion-parameter model, via its API, according to a report by Simon Willison published October 6, 2026. The announcement, titled “Introducing Mistral Large 4: Le chonk,” states the model uses 49 billion active parameters and was trained on the company’s own cluster of 3,800 NVIDIA Grace Blackwell GPUs.

Willison’s report notes that Mistral has promised to release open weights for the model by the “end of this month,” though the current preview is accessible only through the API. The report also indicates the model supports two reasoning levels, though further detail on this feature was not included in the available material.

Willison frames the release as a sign that Mistral, a French AI company, is reasserting itself in the competitive frontier model landscape. The scale of the training infrastructure, nearly 4,000 of NVIDIA’s latest-generation GPUs, points to substantial compute investment behind the model’s development.

No pricing, benchmark comparisons, or specific release date for the open weights version beyond “end of this month” were included in the source material. The eventual open-weights release, if it proceeds as described, would make Mistral Large 4 available for broader use beyond the API preview, a detail Willison’s post identifies as significant for developers tracking open-weight alternatives to closed frontier models.