What Mistral released
Mistral AI has opened the weights of Mistral Large 4, nicknamed Le Chonk. It is a natively multimodal model with 1 trillion parameters. The company says it was trained from scratch on European infrastructure. A public preview ran earlier this month, and Mistral says about three more weeks of reinforcement learning followed before this release.
The model is described as a hybrid of instruction-following and reasoning, built as a mixture-of-experts system with multimodal input, and fluent in more than 160 languages.
Reasoning modes
There are now three modes. Low is for simple tasks and replaces the earlier "None" setting. Medium is for everyday work. High is the default and is meant for the hardest cases that need precision. This gives teams a direct way to trade answer quality against token spend.
Benchmarks, as reported by Mistral
All figures below are Mistral's own presentation of results, including third-party benchmarks it cites.
- Coding: 19.7% on Cognition's private FrontierCode 2.0. On the Artificial Analysis Coding Agent Index it scores 48%, roughly level with GLM-5.3, while using 47% fewer tokens per task on average.
- Terminal work: 41% on Terminal-Bench 4.0, up 13 points from the preview, with 30% fewer output tokens.
- Security: 51% on the Artificial Analysis Cyber Index, and a 96.6% average refusal rate on malicious cyber requests. On the B3 Agent Security Benchmark it resists 97.1% of attacks, against 93.3% in the preview.
- Work tasks: 84.5% on BrowseComp and 73.3% on the Finch finance-workflow benchmark, which Mistral says is second only to Kimi K3 among open models.
Mistral also notes that after the preview it saw some potential for real attack uplift in live traffic and is changing its guardrails. A specialized cybersecurity variant goes only to vetted enterprise design partners under contractual safeguards.
License, pricing and where to get it
The weights are under the Mistral Ecosystem Open-Weight license, a modified MIT license. You may use, modify, fine-tune and commercialize the model. One condition applies: a business that runs a Model-as-a-Service offering needs a separate commercial license from Mistral.
The API price is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Access is through Mistral Studio, the Mistral API and OpenRouter. Weights are on Hugging Face, with deployment via vLLM or SGLang. Microsoft Foundry offers a managed version with private deployment options for Azure developers, and NVIDIA offers hosted endpoints and a NIM container. Mistral Forge is the route for customizing the model on proprietary data.
What is not in the source
The page gives no hardware requirements for self-hosting and no independent verification of the benchmark numbers. Treat the figures as a reason to run your own test, not as a ranking you can rely on.