Why it matters: Open weights let a business or government run the model on infrastructure it controls rather than only through Mistral's API, under a license that allows commercial use but requires a separate deal for companies selling the model as a hosted service.
Mistral AI released the weights of Mistral Large 4 on Oct 29, letting developers download its flagship model from Hugging Face and run it on their own infrastructure. The company first put the model out in public preview on Oct 6 and says the version released this week has had further reinforcement-learning training since then, aimed mainly at coding and defensive cybersecurity.
The model: Mistral describes Mistral Large 4, nicknamed "Le Chonk," as a mixture-of-experts model with 1.05 trillion total parameters and 52 billion active parameters, plus a 1.6 billion-parameter vision encoder. It takes text and images as input, returns text, and has a context window of 1 million tokens. It now offers three reasoning modes, low, medium and high, with high as the default. Hugging Face added support for the model in Transformers v5.22.0 the same day.
Price and license: through Mistral's API the model costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The weights come under the Mistral Ecosystem Open-Weight license, which Mistral calls a modified MIT license. Mistral says the model is free to use, modify, fine-tune and commercialize, except that businesses running a model-as-a-service offering require a separate commercial license from Mistral. It is also offered through Mistral Studio, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry, according to the announcement.
Mistral's benchmark claims: the company says the model scores 48% on the Artificial Analysis Coding Agent Index, which it describes as effectively tied with Z.ai's GLM-5.3 while using 47% fewer tokens per task on average. It reports 41% on Terminal-Bench 4.0, up 13 percentage points from the preview, and 51% on the Artificial Analysis Cyber Index. Mistral also says the model refused malicious cyber requests 96.6% of the time on average across three public test sets. These results are as reported by Mistral; we did not run independent tests.
Security guardrails: Mistral says that while monitoring cybersecurity-related use since the preview, it saw potential for the model to give uplift to real attacks and is adjusting its guardrails. It says a specialized variant for penetration testing and red-teaming is available only to vetted enterprise design partners under contract.
Mistral says a full technical report on the model is coming soon.
Sources
- Mistral AI announcement: Mistral Large 4 open weights, Oct 29, 2026
- Mistral AI documentation: Mistral Large 4 model page
- Hugging Face Transformers release v5.22.0: Mistral Large 4 support, Oct 29, 2026
How we checked this: We read Mistral's Oct 29, 2026 announcement and its Mistral Large 4 model documentation, saved on Oct 30, 2026 at 12:08 PM ET, and checked each figure in this story against them. Benchmark results are Mistral's own reporting.
See something wrong? Tell us and we'll fix it. Every fix is dated and listed on our Corrections page.