Mistral releases Large 4 open weights
Mistral has published the weights of its 1-trillion-parameter Large 4, claiming the top open-weight scores on several coding benchmarks after three more weeks of reinforcement learning since the model's public preview.
Mistral has released the weights of Mistral Large 4, nicknamed "Le Chonk", its natively multimodal flagship with 1 trillion parameters. As we reported earlier this month, the model first launched as a public preview. Mistral says the open release is an improved version, the result of three further weeks of reinforcement learning, and that its gains show "no signs of saturation."
Mistral's headline coding numbers:
- FrontierCode 2.0 (Cognition's private benchmark, which judges whether a maintainer would merge the agent's change): 19.7%, the best score among open-weight models.
- Terminal-Bench 4.0: 41%, up 13 points from the preview, with 30% fewer output tokens.
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task.
In Surge Code's blind pairwise evaluations, testers preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher. The model now offers three reasoning modes (Low, Medium and a default High), with Low replacing the earlier "None" mode.
On security, Large 4 scores 51% on the Artificial Analysis Cyber Index. Mistral reports a 96.6% average refusal rate on malicious cyber requests, which it says is the highest it has measured among open models. Mistral also says its monitoring of live preview traffic showed "a potential for uplift to conduct real attacks," and that it is tightening guardrails as a result. Vetted partners such as Orange Cyberdefense get preview access to a separate variant tuned for penetration testing and red-teaming.
All benchmark figures come from Mistral. The source text is cut off before it covers licensing and deployment details.