The French lab's trillion-parameter open-weight model enters limited preview, with public weights scheduled for October 27
Mistral AI has a new flagship. Its name is Mistral Large 4, and its nickname, officially, is “Le Chonk.”
The French company unveiled the model on October 6, 2026, pitching it as the strongest open-weight AI system built outside China. Mistral also says it outperforms some Chinese rivals. Anyone can judge that claim once the weights go public on October 27.
What Mistral actually built #
Mistral Large 4, or ML4, carries 1 trillion parameters. Only 49 billion of them are active at any given moment.
That gap comes from the model’s design. ML4 is a Mixture-of-Experts system, or MoE. Picture a hospital rather than a single general practitioner. When a patient walks in, they get routed to the right specialists instead of every doctor in the building. An MoE model works the same way, sending each request to a small subset of its internal “experts” rather than firing up the whole network.
ML4 is also natively multimodal. It was built from the ground up to handle more than plain text, rather than having image capabilities bolted on later.
Training took approximately two months. Mistral used about 4,000 NVIDIA Grace Blackwell GPUs housed in its own European data centers.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The benchmark claims, read carefully #
According to Mistral, ML4 scored 38 points on the Artificial Analysis Intelligence Index. The company says that is the highest score for any non-Chinese open-weight model.
MiMo-V2.6-Pro, one Chinese open model, scored 46 on the same index. That puts it comfortably ahead of ML4’s 38.
Mistral is targeting ML4 at specific workloads. These include cybersecurity, finance, coding, manufacturing and visual grounding, which is the ability to connect language to specific elements within an image.
The model supports more than 160 languages. That list covers every official language of the European Union.
Pricing and availability #
For now, ML4 sits in a limited preview. Full access is restricted to select developers and expert users. That changes on October 27, 2026, when Mistral plans to release the model’s weights publicly.
For those who prefer the rental route, Mistral has posted API pricing. Input costs $1.36 per million tokens, and output costs $4.18 per million tokens. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our