Mistral Large 4: France's AI champion bets on open weights and cybersecurity

Mistral AI introduced Mistral Large 4 (ML4), its new flagship model, on October 6. It is multimodal (text and images) and has agent capabilities. For now, it is available as a preview through the API from Mistral. Mistral plans to release its weights on October 27.

Open weights, not quite open source

Companies will be able to download the model and host it on their own infrastructure. However, the training data and method are not published. This is the so-called “open weight” approach, also followed by Meta, DeepSeek and Moonshot AI. According to Mistral, the model was trained for about two months on nearly 4,000 Nvidia Grace Blackwell GPUs (3,800 according to its official blog post) in its European data centers, and it can be deployed with a reduced hardware setup.

It is a mixture of experts model with about 1 trillion parameters in total, of which only a small fraction is used for each word generated. One important point for business use: the license has not been published yet. It will determine whether unrestricted commercial use is possible.

Cybersecurity at the top of the use cases

Mistral makes cybersecurity its first use case. The argument: let companies build their own defenses without depending on closed models, whose access can be restricted. Its CEO Arthur Mensch says ML4 is superior to Chinese models “in some respects, notably cybersecurity.” The company also cites document summarization, coding assistance, 3D design, and analysis of satellite or aerial imagery.

An attempted comeback

In recent months, Mistral's models had dropped out of the top of the main rankings, which are dominated by OpenAI, Anthropic, Google and several Chinese labs. ML4 is the first step of the roadmap funded by the €3 billion funding round (Series D) announced in September. Mistral notably claims the best performance by an open-weight model on Harvey's Legal Agent benchmark. Several results are still pending, however, and the model is expected to keep improving in the coming weeks.

For European organizations subject to GDPR, NIS2 or confidentiality constraints, being able to run a model of this level in-house is a strong argument. We will see on October 27 how it performs in practice.