Mistral Large 4 Puts A 1-Trillion-Parameter Model In Public Preview

Mistral has put a one-trillion-parameter AI model into public preview, but the most consequential part of its pitch is not the number. With Mistral Large 4, the French company wants to offer organisations a powerful model they can eventually run under their own control, instead of depending entirely on a provider’s API. The official announcement on October 6 says developers can try the preview API now. The model weights, which would enable self-deployment, are scheduled for release by the end of the month and are not available yet.

Mistral Large 4, or ML4, is a natively multimodal model built to work across text, images and more complex agentic tasks. Its trillion parameters do not all run for every response. Mistral says 49 billion are active at a time, a design meant to keep a large model more practical to serve. The company trained it on 3,800 Nvidia Grace Blackwell GPUs in its European data centres and is serving the preview from that infrastructure.

The open-weight promise matters because the strongest AI systems are often rented through closed services. That arrangement is convenient, but the provider can change pricing, access, model behaviour or safety rules. A bank, hospital, government agency or software company building a long-term system may want the option to run a model on its own infrastructure and set its own operational policies. As our open-weight AI explainer discusses, downloadable weights can give users that flexibility, although they do not automatically provide the training data or a fully reproducible open-source system.

Mistral is emphasizing this point in cybersecurity. It argues that provider-level refusals can sometimes block legitimate work such as reproducing a vulnerability before fixing it. Self-deployment could let a security team use the model within its own safeguards. That freedom has a difficult other side: a capable model whose weights are widely available can also be adapted or misused beyond the original developer’s control. Mistral says it is red-teaming ML4 with vetted partners and authorities before releasing the weights.

The company presents strong benchmark results for coding, cyber tasks, finance and legal work, including comparisons with both open and closed rivals. Those numbers are useful signals, but readers should treat them as the company’s selected evaluations until more independent testing appears. Real customers will care just as much about reliability, cost, privacy and how the model performs on their own documents and workflows. Mistral lists preview API pricing at $1.36 per million input tokens and $4.18 per million output tokens.

There is also a specifically European business argument. Mistral says ML4 will be available across regions, including a deployment it operates end to end in Europe under European law. For organisations concerned about data location and dependence on a handful of foreign cloud providers, that may be as important as a leaderboard position. The wider question is whether a strong open-weight option can make switching AI suppliers genuinely easier. We will know more when the weights are released, the licence and deployment requirements are clear, and independent users can test ML4 outside Mistral’s own preview.