French AI company Mistral has released the model weights for Mistral Large 4 (ML4), a trillion-parameter AI model designed for coding, cybersecurity, and enterprise applications. The release allows businesses and government agencies to download and deploy the model on their own infrastructure without relying entirely on external AI providers.
According to Mistral’s official announcement on October 29, the model, nicknamed “Le Chonk,” was initially introduced in public preview on October 6. The company has since improved its coding and cybersecurity capabilities through additional training and testing.
Mistral Large 4 Targets Cybersecurity Applications
Mistral says ML4 can help organizations identify software vulnerabilities, investigate security incidents, and respond to threats across their IT infrastructure. By making the model weights available, the company allows organizations to operate the system within their own environments and maintain control over sensitive information.
Pierre Stock, Mistral’s vice president of science, said self-hosting allows businesses and governments to access advanced AI capabilities without depending on continued access from an external provider. He added that organizations could use the model to coordinate responses to major security incidents in real time.
Mistral CEO Arthur Mensch has described AI as a technology that should give organizations greater control rather than increase their dependence on individual providers. The company’s approach differs from closed-model providers such as OpenAI and Anthropic, which primarily distribute their most advanced systems through hosted services.
ML4 Claims Competitive Coding Performance With Fewer Tokens
Mistral says its model delivers performance comparable to Chinese open-weight competitor GLM-5.3 while using 47% fewer tokens per task on average. The comparison is based on Artificial Analysis’ Coding Agent Index, which measures how AI models perform on software development tasks.
The company reported a score of 48% on the benchmark, placing ML4 among the leading open-weight models for agentic coding. Mistral also said the model achieved these results using less computing resources and investment than some closed-model developers.
ML4 contains approximately one trillion parameters, with around 52 billion activated during processing. Its architecture supports text, images, reasoning, and automated workflows, while offering different reasoning settings to balance performance and computational requirements.
Model Weights Available for Self-Hosted Deployment
Mistral has made ML4’s weights available through Hugging Face, alongside access through its application programming interface (API). Organizations can download and customize the model for internal systems, including deployment on their own servers.
Stock also addressed concerns about AI agents acting outside their intended instructions, acknowledging that securing these systems remains challenging. He said the risks should be addressed through technical safeguards rather than preventing organizations from accessing advanced AI capabilities.
Mistral maintains that making defensive AI technology available can help security professionals respond to vulnerabilities, particularly as attackers may also use AI tools. The company has been testing ML4 with cybersecurity organizations, enterprise partners, and government agencies to assess its capabilities in practical environments.
Featured image credits: Mistral AI
For more stories like it, click the +Follow button at the top of this page to follow us.