French lab Mistral put the weights of Mistral Large 4 on Hugging Face on Thursday, three weeks after a preview release. The model’s nickname is Le Chonk.
The numbers are large. About 1T parameters sit in the checkpoint, with 52B active on each step. It is multimodal on the input side and returns both prose and code.
Running it at home is not a laptop job. Documentation for the SGLang inference framework lists eight Nvidia B300 accelerators for the standard build, or four GB300 chips for a compressed variant. Mistral trained the model on 3,800 Grace Blackwell GPUs inside its own European data centers.
The team kept the reinforcement learning loop turning after the preview. Mistral invented training work in house and dialed up the difficulty as the run advanced; the company’s own researchers argue the real constraint is how many tough problems they can still author and check. Vice president of science Pierre Stock described the result as a two-year bet now paying off.
The scoreboard moved. Terminal-Bench 4.0 climbed from 28 to 41 percent, edging past GLM-5.3 at 40. The headline aggregate is less flattering. Artificial Analysis puts the model at 26 on its Intelligence Index, behind open Chinese systems and well short of the leaders from Anthropic and Google.
The release is priced to travel. Through the API, Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens, down from a preview list price of $1.36 and $4.18. Mistral is aiming the release at governments and regulated enterprises that want frontier capability behind their own firewall. That pitch lands hardest in cybersecurity, where sending sensitive traffic to a foreign API is often a non-starter.
