Mistral AI said on 29 October 2026 that the open weights of Mistral Large 4, nicknamed Le Chonk, are now available, so companies can download its most capable model and run it themselves. The model first appeared as a public preview on 6 October, when Forklog reported full weights for 27 October. Mistral says it has since had three more weeks of reinforcement training and customer testing.
“Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors”
Arthur Mensch, co-founder and CEO, Mistral AI, in Mistral's announcement · siliconangle.com ↗
Large 4 is a natively multimodal mixture-of-experts model of about one trillion parameters. The outlets differ on details: SiliconANGLE reports 52 billion active parameters and about 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centres, while Forklog's preview report gave 49 billion active, a 1.05 trillion total and roughly 4,000 GPUs. Forklog also cites a context of up to a million tokens.
Mistral AI artwork from the Large 4 preview announcement
Mistral · Image: Forklog
On performance, Mistral says the model rivals China's GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task, and that its Terminal-Bench 4.0 score rose from 28% to 41%, tying GLM-5.3. Three reasoning modes, low, medium and high, let teams set how deeply it thinks and how many tokens it spends. These are the company's own claims, and Forklog notes that none has been independently checked against Chinese models.
About Mistral Large 4
Mistral Large 4 is a mixture-of-experts model from Mistral AI, announced as a public preview on 6 October 2026 at 1.05 trillion parameters, with open weights promised for the end of that month. The preview was offered through Mistral's API and Mistral Studio, and early access went to developers, cybersecurity specialists and government bodies before the weights were released.
It follows Mistral Large 3, the previous flagship, which came out on 2 December 2025. That model was also a mixture-of-experts design, with 675 billion total parameters and 41 billion active.
Oct 2026
Released
The pitch is control: open weights let firms run the model on their own hardware, behind their own firewalls, without sending sensitive data to a cloud provider. Mistral presents it as a Western-built alternative to huge Chinese coding models from Zhipu and Moonshot AI, aimed at programming, cybersecurity, finance and manufacturing. It is part of a new model generation after the Voxtral 3 voice models, and more models from the family are planned in the coming months.
Sources · 2