French AI lab Mistral said Thursday it released open weights for an improved Mistral Large 4, also called Le Chonk, a 1-trillion-parameter multimodal model it first put in public preview earlier this month¹. The company said extra reinforcement learning raised coding and defensive cybersecurity scores, and it ranked the model as the best open-weight system for coding¹. Wired reported the same pitch: that the refined model matches or outperforms leading Chinese open-weight systems from DeepSeek and Z.ai on coding and cybersecurity².

On company-cited tests, Mistral said Large 4 scored 19.7% on Cognition's FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than a peer it named, and 41% on Terminal-Bench 4.0 after a 13-point gain from the preview¹. It said the model scored 51% on the Artificial Analysis Cyber Index, which it described as measuring vulnerability discovery, reproduction and patching¹. Those figures come from Mistral; Wired did not republish the scoreboard^1,2.

Mistral said the weights can be downloaded from Hugging Face and run with vLLM or SGLang, and that the model is also on its API, OpenRouter and NVIDIA-hosted endpoints¹. It listed API prices of $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens¹. The license is a modified MIT terms set that is free to use, modify and commercialize except that businesses offering the model as a service must obtain a separate commercial license¹.

Pierre Stock, Mistral's vice president of science, told reporters the release was "a pivotal moment in the global AI leadership race"². He said Chinese models could act as threat vectors and compared AI access to electricity that a neighbor should not be able to cut off, arguing open weights cannot be withdrawn once they are public². Mistral said the model was trained from scratch on European infrastructure and pitched it for sovereign, self-hosted use¹.

The company said built-in guardrails try to separate defensive from offensive cyber tasks and that it measured a 96.6% average refusal rate on malicious cyber requests among open models it tested¹. It also said live traffic showed refusals of black-box attacks on live targets, that it still sees potential for real-attack use, and that it is tightening those guardrails¹. A specialized variant for penetration testing is in preview for vetted partners such as Orange Cyberdefense, Mistral said, limited to owned or authorized systems¹.

Rafe Pilling, senior director of threat intelligence at Sophos, told Wired that downloadable models are dual-use and can be abused as well as used productively². He said advanced models of any origin can go rogue without supervision and that monitoring and controls matter more than a nationalist narrative about where a model was trained². Independent third-party confirmation of Mistral's benchmark claims was not in this coverage^1,2.