Mistral Releases Large 4, a 1T-Parameter Multimodal Model, Open Weights by End of Month

On October 6, 2026, Mistral AI released its largest multimodal large model to date, Mistral Large 4 (ML4, nicknamed le Chonk / Le Chonk), and opened a public preview. The model is natively multimodal, with a total of roughly 1 trillion parameters (The Decoder says 1.05 trillion) and about 49 billion activated parameters; the API is already live. There are conflicting accounts about when the weights will be opened: an earlier TechCrunch report said the model was initially only accessible via a public guardrail endpoint, with plans to open the weights within three weeks after completing safety testing; later, The Decoder, Mistral's official announcement, and Ars Technica all said the weights are expected to be available for download at the end of October, and Ars Technica added that anyone can use and customize the model, with the final version to be released before the end of the month. Accounts of training compute also differ: the earlier TechCrunch report said it was trained on just 4,000 Nvidia GPUs; later, reports cited by Simon Willison said it used Mistral's own 3,800 NVIDIA Grace Blackwell GPUs. Officially listed optimized use cases include cybersecurity, finance, and chip design. Ars Technica reported that the model can rival the strongest models from China and the US. Reports also noted that Mistral last month completed a Series D funding round led by Samsung at a €21 billion valuation. The event is still in the model release preview stage, awaiting the opening of weights at the end of the month.

AI overview of reporting · Updated 11 days ago