Mistral Unveils Mistral Large 4, a Trillion-Parameter Open-Weight Model

Mistral AI is back near the front of the pack, or at least much closer to it. The Paris company on Tuesday, October 6, released a public preview of Mistral Large 4, its biggest model to date, which it nicknames “le Chonk.” Developers can try it now through the API on Mistral Studio. Mistral says the weights will be published at the end of the month; Numerama and AFP put the date at October 27.

According to Mistral’s announcement, the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company’s own data centres in Europe, and the preview runs on the same infrastructure. Numerama reports that it is a mixture-of-experts model with about one trillion parameters, 49 billion of them active for each request, up from 675 billion for Mistral Large 3, released in December 2025. AFP adds that it handles text and images and was trained for two months on more than 160 languages.

Mistral is pitching the model hardest on cybersecurity, industry and finance. Co-founder Guillaume Lample said the aim is to allow “companies and governments to defend themselves against malicious actors” who get around the safeguards of closed models to launch cyberattacks (translated from French), AFP reported. Until the weights ship, Numerama says, Mistral is offering a less restricted version with extended cybersecurity capabilities to developers, security officials and state authorities.

What it means in practice

Lample was candid about the remaining distance. Speaking to Le Figaro, as quoted by 01net, he called Large 4 “a model that does not completely close the gap with frontier models, but reduces it significantly” (translated from French). For French buyers, the pitch is less about leaderboard rank than about control: open weights let ministries and companies run the model on infrastructure they choose, under European law.

Large 4 is the first milestone on the roadmap funded by Mistral’s €3 billion Series D, which AFP says valued the company at more than €21 billion. It is also a test of the compute push behind France’s sovereign AI campus projects, and it arrives as Inria builds out an AI evaluation and safety programme that could put such claims to the test.

What to watch: the weights and full benchmarks promised for late October, independent testing of the less restricted cyber version, and the specialised industry models Mistral says it will build on top of Large 4.