Mistral AI has released the full weights of Mistral Large 4, a 1-trillion-parameter multimodal model the company nicknamed "Le Chonk." Anyone can download it and run it on their own hardware. Mistral calls it the best open-weight coding model: 48% on the Artificial Analysis Coding Agent Index, tied with the leading open rival while using 47% fewer tokens per task. In blind human comparisons of code, only Anthropic's Claude Opus 5.5 beat it.

The headline feature is cybersecurity, and that is where it gets awkward. It scores 51% on a benchmark for finding and patching vulnerabilities, and refuses 96.6% of malicious cyber requests. Yet Mistral admits that after watching live traffic during the preview, it sees "potential for uplift" for real attacks, and it is tightening the guardrails. Guardrails on weights anyone can already download. Mistral's bet is that defenders need a model they control more than attackers need one more tool, and the EU cyber agency ENISA and CrowdStrike are already testing it.

The model ships under a modified MIT license: free to use, fine-tune and sell, unless you run a model-as-a-service business, which needs a separate commercial deal. API pricing is $0.70 per million input tokens and $2.10 per million output. Every benchmark above comes from Mistral itself, so independent tests will matter.

Why it matters: Europe now has a frontier-scale open model trained from scratch at home, giving banks, governments and security teams a sovereign alternative to someone else's API, and a live test of whether "open" and "safe for cyber" can survive in the same release.

Source: MistralAI