Mistral AI releases Large 4, a 1-trillion-parameter open-weight model with a cybersecurity focus

French AI startup Mistral AI publicly released its 1-trillion-parameter Large 4 model on Thursday, letting companies and developers run advanced AI on their own infrastructure without relying on cloud services controlled by U.S. technology giants. The release puts cybersecurity at the center of Mistral’s effort to challenge proprietary AI systems from OpenAI, Google, and Anthropic.

The October 29 release marks a significant step for the Paris-based startup, which first introduced Large 4 in a public preview on October 6. Since then, Mistral spent the intervening weeks working with government agencies and corporate customers to test the model and improve its performance before making it publicly available.

Unlike proprietary models that primarily run through their developers’ platforms, Large 4 lets organizations download its trained weights and deploy the software on infrastructure they control. That distinction matters for banks, government agencies, and cybersecurity teams handling sensitive information or operating systems that require uninterrupted access.

Mistral says Large 4 is its most capable openly available model for cybersecurity, with the ability to help security teams discover software vulnerabilities, investigate threats, and develop fixes.

The model uses a mixture-of-experts architecture with approximately 1 trillion parameters, activating only a fraction during each operation. It supports coding, multimodal inputs, autonomous AI workflows, and a context window of up to one million tokens. The architecture is intended to deliver advanced capabilities without activating the entire model for every request.

Mistral claims Large 4 ranks among the world’s strongest AI models for cybersecurity. In its October announcement, the company reported an 82% score on a vulnerability reproduction and patching benchmark and said the model solved 93% of challenges in Cybench, a cybersecurity evaluation based on security competitions.

Those results suggest strong potential, but broader independent testing of the final public release will be needed to establish how it compares with competing systems in real-world deployments.

Mistral challenges closed AI models as Europe pushes for AI independence

The release comes amid a growing debate over who should control access to advanced AI models, particularly systems that can identify and exploit software vulnerabilities.

In an October 6 interview with WIRED, Mistral cofounder and chief scientist Guillaume Lample argued that relying on proprietary models creates risks for companies that depend on continued access.

“If you use a closed model, there is no guarantee it will still be there tomorrow,” Lample told WIRED.

That concern carries particular weight in cybersecurity, where restrictions imposed by an AI provider could interrupt legitimate security investigations or incident response efforts.

Mistral Vice President of Science Pierre Stock made a similar argument in comments Reuters reported on Thursday.

“I’m not comfortable living in a world in which open weights do not exist,” Stock told Reuters.

Stock acknowledged a fundamental tradeoff. Once a model’s weights become publicly available, “it’s impossible to contain access.” Organizations gain independence from providers, but malicious actors can obtain the same technology.

The release is already attracting commercial interest. Reuters reported Thursday that Orange Cyberdefense, the cybersecurity division of French telecommunications group Orange, has partnered with Mistral to integrate a cybersecurity-focused version of its latest model into selected services.

The partnership will initially focus on identifying weaknesses in customers’ defenses and speeding up investigations after security incidents. Orange plans to run the models on computing infrastructure in Europe.

For Mistral, the stakes extend beyond cybersecurity. Chinese AI developers, including DeepSeek and Alibaba, have become influential suppliers of open-weight models, leaving European companies competing against both China’s open-model ecosystem and America’s largest proprietary AI providers.

The company raised €3 billion in September and has invested in European computing infrastructure to train and deploy its models. Mistral trained Large 4 on approximately 3,800 Nvidia Grace Blackwell GPUs in Mistral-operated European data centers.

Founded in 2023 by former DeepMind and Meta researchers Arthur Mensch, Guillaume Lample, and Timothée Lacroix, Mistral has positioned itself as Europe’s leading independent AI challenger.

The bigger question is whether organizations will increasingly prioritize control over their AI systems rather than relying exclusively on the most capable proprietary models.

With Large 4 now publicly available, Mistral is betting that ownership, deployment freedom, and independence from outside providers could become just as important as benchmark performance in the next phase of AI competition.

Mistral AI Team