Mistral Opens Weights of Trillion-Parameter Large 4: 'Le Chonk' Can Be Downloaded and Self-Hosted, but Reselling It as a Service Needs a Separate Licence

French AI lab Mistral AI published the full weights of its flagship model Large 4, nicknamed "Le Chonk," on October 29. The model first appeared in a public preview on October 6 and went open-weight after about three more weeks of reinforcement learning and customer testing; the weights are on Hugging Face and the model is also served through Mistral's own API, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry.

A trillion parameters and a million-token window, built for coding and cyber defence

Large 4 is a natively multimodal mixture-of-experts model: about 1 trillion parameters in total, roughly 52 billion active per token (some reports put it at 49 billion), reading text and images and producing text and code, with a 1-million-token context window and, according to its model card, support for more than 160 languages. Mistral positions it for coding and cyber defence: the company says its Terminal-Bench 4.0 score rose from 28% at preview to 41%, and that it scores 48% on Artificial Analysis' Coding Agent Index, tied with open-weight leader GLM-5.3 while using 47% fewer tokens per task — company and benchmark-maker figures, not independent everyday tests.

Open, with one boundary: reselling the model as a service needs a separate licence

Large 4 ships under Mistral's Ecosystem Open-Weight Licence, a modified MIT licence: anyone may use, modify, fine-tune and sell products built on it, with one exception — a business offering the model itself as a paid hosted service (Model-as-a-Service) must first buy a separate commercial licence from Mistral. Internal use and access through Mistral's official products are not affected. Hosted pricing is about $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens, roughly half the preview price.

Self-hosting is heavy, and independent scores are mixed

Running Large 4 yourself is a serious infrastructure project: inference documentation puts the standard version at eight NVIDIA B300 GPUs, and even the compressed variant at four GB300s. On independent scoring, Artificial Analysis' Intelligence Index v5 rates Large 4 at 26 points — the highest among Western open-weight models, but only fifth among open weights overall, behind DeepSeek V4.1 Flash and GLM-5.3 at 32. Mistral also says the model refuses 96.6% of malicious cyber requests on average, and that a penetration-testing version is available only to selected partners after review and under contract.