Mistral has released the downloadable weights for Mistral Large 4, giving developers and companies the option to run the model on their own infrastructure instead of relying only on a hosted service. The model, also known as Le Chonk, was first introduced in preview earlier this month. The updated weights became available on October 29, following several weeks of additional training. Reuters reported the public release on Thursday, while Mistral outlined the changes in its announcement.
Large 4 is a multimodal model with around one trillion parameters. Mistral says the updated version is stronger at coding and reasoning and is designed to help security teams examine source code, reproduce vulnerabilities in isolated environments and test fixes. The company reports a 51% score on the Artificial Analysis Cyber Index, which measures how well a model can discover, validate and patch software vulnerabilities. That figure is a company-reported benchmark result, not a guarantee of performance in every security environment.
The release also highlights the risks that come with making powerful security tools widely available. Mistral says its monitoring showed the model refusing some attempts to attack live systems, but the company still saw the potential for it to increase the ability to carry out real-world attacks. It says it is adjusting the model’s safeguards and will continue updating them. Mistral reports a 96.6% refusal rate for malicious cyber requests in its own testing, although that result is based on its selected test prompts and does not remove the possibility of misuse.
The concern comes as AI-powered cyberattacks are drawing greater attention from security teams. A model that can help defenders reproduce and fix vulnerabilities may also give attackers another tool to explore weaknesses, particularly when downloadable weights can be modified and run without a central provider controlling access.
The weights are available on Hugging Face under Mistral’s modified MIT-style open-weight license. Developers can download, modify, fine-tune and commercialize the model, but businesses offering it as a model-as-a-service must obtain a separate commercial license from Mistral. The model is also available through Mistral’s API and Microsoft Foundry.
Mistral Releases Large 4 Model With Open Weights and Cybersecurity Focus



