Mistral has announced Mistral Large 4, or “Le Chonk”, its new flagship model, and says it is available now via API in public preview. The company also said it plans to release the model’s open weights at the end of October.
According to Mistral and co-founder Guillaume Lample, Mistral Large 4 is a 1 trillion-parameter model with 49 billion active parameters that was trained natively with multimodal capabilities. Lample also said the model was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters.
That scale matters partly because Mistral is pitching Large 4 as more than another very large model with a memorable nickname. In its launch post, the company said the model is “the best open weights model from US or Europe on aggregated benchmarks” and described it as state of the art on workloads including cyber defense, manufacturing, and finance. Mistral also said it surpasses closed frontier models on visual grounding.
The pitch is as much about cybersecurity as scale
Mistral’s own follow-up posts lean especially hard on cybersecurity. The company said Mistral Large 4 is “one of the world's strongest AI models for cybersecurity”, and another post claimed the model solved 18 of 19 challenges in a CTF speedrun.
That emphasis also shows up in third-party benchmark commentary. Artificial Analysis said Mistral Large 4 is in research public preview, repeated Mistral’s plan to release weights at the end of October, and said the model scored 50 on its Cyber Index, level with GLM-5.3-Flash. Artificial Analysis also said the model scored 38 on its Intelligence Index.
Those numbers help explain the company’s angle. Open-weight model buyers usually are not just shopping for a chatbot leaderboard winner; they are often looking for a model they can host themselves, audit, and tune for specific enterprise use cases. Mistral’s launch messaging repeatedly frames Large 4 around that kind of practical deployment, especially for security-sensitive work.
