Mistral releases Mistral Large 4 as open weights

On October 6, Mistral put Large 4 — nicknamed Le Chonk — into public preview through its API, and promised the weights by the end of the month. On October 29, it released those weights, for a model it says improved during three more weeks of reinforcement learning and feedback from developers, early-access customers, and partners.

The release changes what Large 4 is: not just a hosted model, but weights you can download from Hugging Face and serve yourself with vLLM or SGLang. It is also available through Mistral Studio and the Mistral API, on OpenRouter, through Microsoft Foundry, and as an NVIDIA NIM container. According to Mistral's model card and documentation, the model is a natively multimodal mixture of experts: 1.05 trillion parameters in total (Mistral's announcement rounds this to 1 trillion), 49 billion active per token (52 billion counting embeddings and output layers), a 1.6-billion-parameter vision encoder, and a 1-million-token context window. Mistral says it was trained from scratch in Europe, on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers.

The benchmark figures below are as reported by Mistral. Several are scores on third-party indexes run by Artificial Analysis, and Hugging Face's launch post gives slightly different figures for some, such as 49.8 percent on the Coding Agent Index against the 48 percent in Mistral's announcement. Mistral says a full technical report is coming soon.

On Cognition's FrontierCode 2.0, which judges whether a repository maintainer would merge an agent's change, Mistral reports 19.7 percent, the lead among open-weight models. On the Artificial Analysis Coding Agent Index it reports 48 percent, effectively tied with the leading open-weight model, GLM-5.3, while using 47 percent fewer tokens per task on average. On Terminal-Bench 4.0 it reports 41 percent, up 13 points from the preview, with output tokens down 30 percent.

Three reasoning modes now let a developer trade cost against precision: Low for simple tasks, Medium for everyday work, and High, the default, for the hardest cases. In Surge Code's blind pairwise tests of human preference on code, Mistral says Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, and ranked behind only Claude Opus 5.5 among the models evaluated. Through the API, Mistral now prices it at $0.70 per million input tokens, $0.07 cached, and $2.10 output.

The other half of the release is defensive cybersecurity. Mistral reports 51 percent on the Artificial Analysis Cyber Index, which covers finding a vulnerability, reproducing and validating it, and patching it without breaking existing functions. Its guardrails, Mistral says, tell defensive work from malicious requests: the model refused an average of 96.6 percent of malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm, the highest rate Mistral measured among open models.

Mistral also describes a risk. It says monitoring of live traffic showed the model refusing black-box attacks on live targets, but also a potential for uplift to conduct real attacks, and that it is modifying its guardrails in response. A specialized variant tuned for penetration testing, vulnerability assessment, and red-teaming is going only to vetted enterprise design partners such as Orange Cyberdefense, restricted to owned or explicitly authorized systems. Mistral vice president of science Pierre Stock told Reuters, as reported by PYMNTS, that once the model is released it is “impossible to contain access.”

Mistral describes Large 4 as its largest and most capable model to date. With the weights, organizations can run it on hardware they control, for example for coding agents, for defensive security work on their own systems, or to fine-tune it on their own data. Mistral says that open weights mean access to the model cannot be cut off by an API provider.

Self-hosting needs data-center GPUs: the serving examples on Mistral's model cards use four to eight GPUs. The model is also available as a hosted service through Mistral's API, OpenRouter, and Microsoft Foundry. It is released under the Mistral Ecosystem Open-Weight License, a modified MIT license: free to use, modify, fine-tune, and commercialize, except that a business operating a model-as-a-service offering must obtain a separate commercial license from Mistral.