France strikes back: Mistral gives away its trillion-parameter AI for free

Mistral has released the open weights of Mistral Large 4, which the company calls Le Chonk. The nickname comes from internet slang for chubby cats, and it fits: the model has 1 trillion parameters. It is multimodal and, as Mistral stresses, trained from scratch on European infrastructure.

A public preview came out earlier in October. Since then the model got three more weeks of reinforcement learning, mostly on code and cybersecurity. According to Mistral, it is now the best open-weight model for coding. On the private FrontierCode 2.0 benchmark, which checks whether a repo maintainer would merge the agent's change, it leads with 19.7%. On the Artificial Analysis Coding Agent Index it scores 48%, tied with the top open model GLM-5.3 while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it jumped 13 points to 41%.

In blind human comparisons on code, people preferred Le Chonk's answers over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only the closed Claude Opus 5.5 ranked higher. In one demo the model, with no human help, built a live 3D map of about 19,000 satellites and debris pieces in orbit: it pulled the data, wrote the app and fixed its own bugs.

The second bet is security. On the Cyber Index it scores 51%, finding vulnerabilities, reproducing them and patching them without breaking anything else. At the same time it refuses 96.6% of clearly malicious hacking requests. There are now three reasoning modes, from quick to very thorough.

The weights are on Hugging Face and run locally with vLLM or SGLang. The license is a modified MIT: free to use, fine-tune and make money with, and only companies selling the model as a hosted service need a separate license. Through the API it costs $0.70 per million input tokens and $2.10 per million output tokens.

For a long time Chinese labs were the only ones giving away models this big. Now they have a rival from Europe, and it is gunning for code.