Mistral Releases Large 4, an Open Model With a Trillion Parameters

Mistral, the French AI lab, has released fully open weights for Large 4, code-named Le Chonk. It is a 1-trillion-parameter model trained from scratch on European infrastructure, and it supports more than 160 languages. On the Artificial Analysis Coding Agent Index, a benchmark for coding agents, it scored 48%. That is nearly on par with the best open model, GLM-5.3, but it uses 47% fewer tokens per task. In a test measuring refusals of malicious cybersecurity requests, it refused 96.6% of the time. Mistral says that is the highest score it has measured among open models. API access costs $0.70 per million input tokens and $2.10 per million output tokens.

Saving tokens and refusing malicious requests at a high rate are exactly what matters in real agent work, not just on a benchmark chart. But the model has a trillion parameters. Running it outside Mistral's own cloud takes GPUs worth hundreds of thousands of dollars, so for most people this stays a cloud service rather than a way to skip a subscription. On the hardest agentic coding tasks, Opus and Fable are still ahead. For a narrower job, one where cybersecurity work and data need to stay in Europe, this is a solid second-tier open model, without the usual marketing overreach.