# Mistral released the open weights of an improved Mistral Large 4, and an independent test ranks it the best open model built outside China.

> Mistral released the open weights of an improved Mistral Large 4. An independent test ranks it the best open model built outside China.

Published 2026-10-29T12:20:21-07:00 · News · Canonical: https://news.littlewolflabs.com/posts/2026-10-29-mistral-large-4-open-weights

Mistral on Thursday released the weights of an updated Mistral Large 4, nicknamed "Le Chonk," which it first put out as a public preview earlier this month ([Mistral](https://mistral.ai/news/mistral-large-4-oss/), [Mistral on X](https://x.com/MistralAI/status/2115811242556137820)).



Open weights means you can download the model and run or customize it yourself instead of only using it through Mistral's service. It's a very large mixture-of-experts model: about 1 trillion parameters in total, with roughly 52 billion active for each word it generates, plus a 1 million-token context window and image input. Mistral says it's "the best open-weight model for coding," with leading scores on Cognition's private FrontierCode 2.0 test and on Terminal-Bench 4.0, and strong results on defensive cybersecurity work. Those are Mistral's claims. Artificial Analysis, which tested it independently, calls it "narrowly" the #1 open-weight model developed outside China, one point ahead of Muse Spark Open on its Intelligence Index.



It still trails the leading open models from China, DeepSeek V4.1 Flash and GLM-5.3, by six points, and in its runs scored 15% on FrontierCode 2.0 against 29% and 22% for those two ([Artificial Analysis on X](https://x.com/ArtificialAnlys/status/2115820547103952925)). Artificial Analysis also found it uses far fewer tokens per task than the preview and hallucinates less (28%, down from 41%). The API price is $0.70 per million input tokens and $2.10 per million output tokens, about half the preview's price. It's available in Mistral Studio, the Mistral API, OpenRouter and Mistral's Vibe Code tool, and the weights are on [Hugging Face](https://huggingface.co/blog/mistral-large4). The license is free to use, modify and sell products with, but companies that sell access to the model itself as a hosted service need a separate commercial license from Mistral. NVIDIA says it was trained from scratch on 3,800 of its Grace Blackwell GPUs ([NVIDIA AI on X](https://x.com/NVIDIAAI/status/2115817617911157099)).

*What to weigh:* this is a strong, cheap option, especially if you want a model that isn't from a US or Chinese lab. But running it yourself takes multiple data-center GPUs, so most people will use it through the API or OpenRouter. And independent tests put it a step behind the top Chinese open models on coding, despite Mistral's benchmark claims.