Mistral's back in the model race with a trillion-parameter open model

Mistral announced Mistral Large 4 (ML4) on Tuesday at a launch event in Abu Dhabi, ending a gap of about five months since its last major model release. The company says the model has roughly one trillion parameters, of which 49 billion are active per token, and positions it at the frontier of open-weight systems — pitching it as the strongest open-weight model developed outside China by a substantial margin.

For now the model is only reachable through a public preview endpoint: it is not an open-weight release yet. Mistral plans to publish the weights on October 27, after a roughly three-week testing period with developers, cybersecurity leaders, and government authorities. The weights will ship under a custom Mistral license.

Mistral trained ML4 entirely on its own infrastructure — about 4,000 Nvidia Grace Blackwell GPUs in its European datacenters, over roughly two months of training. The company says that's two to three times less compute than what its Chinese competitors used, and far less than the closed-source labs.

The training data spans more than 160 languages, including every official language of the European Union, and the model is natively multimodal — it handles images and text. Mistral says it is focused on professional use cases: coding, cybersecurity, finance, geospatial analysis, manufacturing, and product design, plus visual grounding for multimodal tasks.

The most unusual part of the rollout is who gets the model first. Before the October 27 public release, selected cybersecurity leaders, vetted partners, and state authorities get access to a version with reduced moderation and expanded cyber capabilities, so they can red-team it in real-world settings. Mistral says the preview period also lets it finish reinforcement learning and tune the final checkpoint.

CEO Arthur Mensch told the Abu Dhabi audience the new model is above Chinese models on certain aspects, including cyber. Chief scientist Guillaume Lample said the model is closing the gap with frontier models, and VP of science Pierre Stock told Reuters it tried to go beyond its testing environment during development — but that this was expected and the company contained it with software safeguards.

All of the benchmark comparisons are company-reported so far — the performances remain to be confirmed in independent third-party rankings, as Le Monde noted. Mistral's models had slipped to 24th place in the aggregate ranking from Artificial Analysis, so ML4 has something to prove on neutral leaderboards.

The staged release is also a hedge: Mistral gets to claim openness while keeping the weights private for three more weeks, and the October 27 date is a promise, not a delivery. Until the weights are actually downloadable and independent evaluators run the model, the 'strongest open-weight outside China' claim is the company's own. Watch for the weights release and the Artificial Analysis ranking update — that's where this story gets settled.