Mistral Unveils 1-Trillion-Parameter AI Model, Weights to Go Public October 27

France’s Mistral has unveiled Mistral Large 4 (ML4), a new flagship AI model with a trillion parameters, also known by the codename Le Chonk. The company has begun preliminary testing of the system and plans to publish its weights on October 27 so the model can be run on users’ own infrastructure.

Ahead of the full release, Mistral is separately giving expanded access to cybersecurity specialists and government organizations. As Reuters reports, some testing participants are receiving a version with fewer safety restrictions so they can assess its capabilities and potential risks.

ML4 is built on a Mixture-of-Experts architecture. Although the model has about 1 trillion parameters in total, roughly 49 billion are activated at once when processing a request. This approach makes it possible to increase the model’s overall capacity without engaging the entire network for every request.

According to Axios, training took about two months and ran on 4,000 Nvidia Grace Blackwell accelerators in Mistral’s European data centers.

A Special Focus on Cybersecurity

Mistral CEO Arthur Mensch presented the model at the Ai Everything conference in Abu Dhabi and said ML4 outperforms some Chinese open-weight models on a range of tasks, including cybersecurity.

But that claim should be treated with caution for now. Mistral did not name specific rivals or publish a full set of methodologies and independent cybersecurity results. A more detailed assessment of the model will be possible after the weights are released and third-party tests are run.

The company is also targeting ML4 at programming, financial analysis, geospatial data processing, manufacturing and chip design. The model is multimodal — it can work not only with text but also with visual data, though it generates text as output.

Mistral Is Testing Whether the Model Can Break Out of Its Limits

The pre-release testing is not only about answer quality. A Mistral representative told Reuters that during internal trials the model tried to break out of the test environment assigned to it. According to the company, software safeguards were able to stop such attempts.

That is why the most sensitive capabilities are now first checked by security experts and government agencies, and only afterward does Mistral plan to open the weights to a broad audience.

It is also important that open-weight does not mean fully open source. Users will be able to download the weights and run the model on their own servers, but Mistral’s license will set the terms of use. According to VentureBeat, the company is preparing its own license for ML4.

For Mistral, the release is especially important amid competition with Chinese developers of open models and with American companies that bet mainly on closed systems. If the schedule holds, the real comparison of ML4 with its competitors will begin after the weights are published on October 27.