Mistral AI Launches ML4 with Open Weights, Cybersecurity Priority

Mistral AI unveils ML4, a multimodal and agentic model trained for two months on 4,000 Nvidia Grace Blackwell GPUs in Europe. The weights are set to be released on October 27, and the model will be offered in open-weight, with options for sovereign hosting. The startup, recently endowed with €3 billion, is particularly targeting cybersecurity and claims high performance, while noting that several benchmarks are still ongoing.

4,000 GPUs today, aiming for one gigawatt by 2030

ML4 was trained for two months on 4,000 Nvidia Grace Blackwell GPUs, installed in Mistral's European data centers. Guillaume Lample specifies that the model was designed to be deployable with minimal configuration. The company has its own infrastructure in Europe and plans to significantly increase its capabilities, aiming to reach one gigawatt of power by 2030. Mistral wants to demonstrate that it is possible to stay close to the technological frontier with more limited computing resources, even though its current means are still below those of American players like OpenAI or Anthropic.

Claimed performance and incomplete evaluations

Mistral highlights high performance for ML4 across various sectors. In the legal field, the company claims to have achieved the highest score to date for an open-weight model on Harvey's Legal Agent benchmark. It also emphasizes performance in finance, industry, and specific multimodal tasks. However, Mistral notes that it is still awaiting final results from several benchmarks and announces a new learning phase in the coming weeks. Arthur Mensch stated at a conference in Abu Dhabi that ML4 outperforms Chinese models in certain aspects, particularly in cybersecurity. The confirmation of ML4's positioning will depend on upcoming results.

Sovereignty of deployments and controlled openness

Companies will be able to host ML4 on their own infrastructure or access it via the API in their chosen region. Mistral emphasizes a sovereign European region, where data remains under the jurisdiction of the European Union, and states that the model is designed to be customized according to users' data, tools, and security requirements. ML4 will be offered in open-weight: clients will be able to retrieve the weights for hosting, without the entire training data or methodology being made public. Sovereignty is presented as a central argument by Mistral.

Cybersecurity in focus, but broader use cases

Cybersecurity is among the main applications highlighted by Mistral for ML4. The company believes that many organizations remain vulnerable and that access to closed models can pose challenges, hence the interest in a powerful and reusable model without limits. Recent work has shown rapid progress in AI agents for detecting security vulnerabilities and automating certain tasks. Beyond security issues, ML4 is described as a solution for creating 3D objects for industry, synthesizing large amounts of documents, assisting in software development, and recognizing objects in images from satellites or aerial photography. Mistral also mentions applications for damage estimation by insurance companies, threat detection along high-voltage power lines, and monitoring the condition of agricultural crops.

Availability, funding, and the challenge of catching up

ML4 is available in preview via Mistral's API, with the release of weights scheduled for October 27. This multimodal model, featuring agentic capabilities and nicknamed "Le Chonk," is part of a trend followed by Meta, DeepSeek, and Moonshot AI, which offer open-weight models. In September, Mistral raised €3 billion in a Series D funding round aimed at increasing its computing capabilities and accelerating research. ML4 is presented as the first step in this roadmap, with goals for research, data center expansion, and commercial and international growth. The stated challenge is to close the gap with the best American and Chinese models, as recent rankings have been dominated by OpenAI, Anthropic, Google, and Chinese labs. Mistral aims to get closer to the global frontier. Arthur Mensch used a video of Floria Gueï's relay at the 2014 European Championships, accompanied by the word "exciting," to announce this strategy.