Mistral Large 4: Mistral unveils ‘the Chonk’, an open model with a trillion parameters, tailored for agents

Mistral has launched a preview of its new flagship model, trained on 4,000 Grace Blackwell GPUs. The results will be published on 27 October, with a focus on cyber security.

Mistral AI opened the public preview of Mistral Large 4, nicknamed "the Chonk" internally, on Tuesday. This new flagship model has a trillion parameters, its weights will be freely downloadable and it is primarily aimed at generalist agentic uses. The Parisian laboratory claims that it competes against the best closed models while holding its own against open models three times larger, and that it has been trained on 4,000 Nvidia Grace Blackwell GPUs for a significantly lower cost than its competitors.

Cybersecurity as an argument

Mistral insists on the model's performance in both attack and computer defense. For Guillaume Lample, co-founder and scientific director, the opening of the weights is decisive here: a company that builds its defenses on a model it owns does not have to fear that its supplier will remove it or restrain it overnight, even though these uses consume a lot of inference. The model also claims good results in code, specialized work and multimodality.

With its size, Mistral Large 4 will not run on a personal computer and will remain out of reach for many universities: it mainly targets on-premise or private cloud deployments, in addition to Mistral's API. The weights will be published on October 27. The launch offers a European alternative in a segment of open models currently dominated by Chinese laboratories.