Mistral AI to Release a Large Model Within Weeks, With Two More Generations Within a Year

Mistral AI will release a large AI model within a matter of weeks. CEO Arthur Mensch laid out that timeline in an interview with Le Monde published on September 24. He also said the two generations that follow would arrive within the next 12 months. This was not a product launch event but an interview in which he described the company’s development and release plans.

What will be different about the new model

Mensch explained that the computing resources used to train models will increase roughly 20-fold. That is a plan to scale up the resources devoted to training. It does not mean the new model will be 20 times better. Scaling up resources and actually improving performance are two separate things.

Mistral AI had earlier announced, on September 8, that it had raised 3 billion euros. The round put its valuation at more than 21 billion euros. Following that funding, this interview set out the release plan for a large model and its successor generations. Announcing a plan, however, is a different stage from users actually getting their hands on the product.

Why “you can run it yourself” is the pitch

The strength Mistral AI keeps emphasizing is that companies can run its models on their own servers. Mistral Large 3, released in December 2025, made the trained model itself available under Apache 2.0 terms. That is a form of release under which anyone is free to use and modify it. Mensch pointed to this release approach, combined with companies running the models on their own infrastructure, as what sets his company apart.

For enterprise users, where AI runs and how freely it can be modified are important criteria in choosing a system. Mistral Large 3 gave companies the option to deploy and modify it themselves. For the models still to come, the question is not only performance but whether this same mode of use carries over.

What is the basis for saying AI can be controlled

Mensch said AI is software and therefore can be controlled. As methods, he cited behavioral testing before release, running models in environments isolated from other systems, and monitoring by a separate model. He also mentioned the process of evaluating models in real working conditions. At the same time, he acknowledged that behavior is hard to predict, which makes development and verification complex.

This is a methodology for managing AI. It is not a result showing that the forthcoming model has passed verification. Mensch also argued that some US AI companies invoke safety risks in order to block competition from openly released models. That is his own reading of rivals’ intentions. His remarks alone do not prove what those companies are really thinking, nor do they establish that AI can be fully controlled.

What Korean users can verify right now is language support in the existing models. Mistral AI’s official documentation lists Korean among the languages supported by its current language models. That should not be stretched into an assumption about how well the forthcoming model will handle Korean. Anyone planning to use it for work in Korea will need to assess the new model’s Korean-language capability separately.

The core of this announcement is the release plan for successor models and an emphasis on AI that companies can run on their own servers. What actually changes for users depends on the performance and terms of use of the real product. At launch, it is worth checking pricing, availability in Korea, and Korean-language performance together. It will also be worth watching how the control methods Mensch described hold up in actual testing.