Mistral Large 4: a preview of a trillion-parameter model

On October 6, 2026, Mistral AI opened a public preview of Mistral Large 4. It is the largest and most capable model in the company's history: a trillion parameters with 52 billion active, native multimodality, and an emphasis on code, agentic scenarios and cybersecurity. The preview API can be tried out in Mistral Studio.

The announcement on the blog does not begin with benchmarks. "Unofficially ML4, entirely officially le Chonk," Mistral writes, and the cat nickname, which can be translated as "chonk", ends up right in the title of the post. The company clearly decided not to pretend it was talking about something more respectable.

According to its data, ML4 is already on par with the strongest open models in the world and noticeably ahead of open models created in the US or Europe. On enterprise tasks in cybersecurity, finance and law it is called the best among open solutions, and in object recognition in images it slightly edges out the closed flagship: 42% versus 41% in the Dense 200 test.

Cybersecurity as a separate argument

Special emphasis is placed on security. According to the estimates Mistral cites, ML4 is among the world's top-5 models for finding and fixing vulnerabilities in real code and leads by a wide margin among open models created outside China. In a test that requires reproducing a vulnerability in open-source software and then closing it, the result was 82%, and in the Cybench set it solved 93% of 40 tasks. The company notes that some closed models show close to zero on the same tasks because they refuse to perform them, while defenders need a tool that proves a vulnerability is real.

Own capacity in Europe

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell accelerators in Mistral's own data centers in Europe, where the preview also runs. A notable share of the training data was multilingual: more than 160 languages, including all official languages of the European Union. ML4 is described as the first stage of a roadmap funded by a Series D round of 3 billion euros, the largest in the history of European technology companies.

Weights and what comes next

The weights are promised to be released by the end of October. Until then, the model is undergoing red-teaming in real-world conditions with cybersecurity leaders, vetted partners and government agencies: they receive the same model with weakened moderation and expanded cyber capabilities. Along with the weights, architecture details, additional benchmarks and a description of the post-training are promised. ML4 will become the foundation for a new generation of specialized models.

Image:

Translation editor: Салтанов А.