Mistral has published the weights of Mistral Large 4, the 1-trillion-parameter multimodal model it calls Le Chonk. The release follows a public preview earlier this month and lands alongside a claim that it now ranks as the best open-weight model for coding. Mistral says the model was trained from scratch on European infrastructure, giving organisations a documented lineage that runs from the first pretraining step through to the final release.
On Cognition's FrontierCode benchmark, Le Chonk takes the top spot among open models. Artificial Analysis places it narrowly first among open-weight models built outside China, with an Intelligence Index score of 26, one point above the preview build. The larger change is efficiency: it burns roughly 82,000 output tokens per index task, about 39 percent fewer than the preview, and its hallucination rate fell from 41 percent to 28 percent.
Pricing dropped by roughly half to 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, with cached input at 0.07 dollars. The model ships with three reasoning modes, a 1M token context window, and text plus image input. It is downloadable on Hugging Face under the MEOW license, a Mistral open-weight licence that still requires a separate commercial agreement from anyone running it as a model-as-a-service business.