Mistral Unveils ML4 Weights for Cybersecurity

The French publisher has made the weights of Mistral Large 4 available for free, a model with one trillion parameters introduced earlier this month and nicknamed "The Chonk." The company aims for on-site deployments by public and private actors for cyber defense, claiming efficiency gains over GLM-5.3 as well as benchmark results in agentic coding. The weights are available on Hugging Face and via API.

Mistral Highlights Control and "Sovereign" Cyber Defense

Mistral asserts that ML4 provides sovereign cyber defense capabilities to businesses and administrations, allowing both private and public institutions to download, deploy, and own cutting-edge AI to detect, analyze, and respond to threats in code and IT infrastructure. According to Pierre Stock, the release of the weights gives organizations more control over how to use AI for defense, with access that cannot be cut off due to the open-weight nature of the model. He states that the model can be used to identify, coordinate, and respond in real-time to major security incidents, and that the openness allows for on-site deployment within the corporate ecosystem. He adds that advancements in AI for cybersecurity are already helping professionals detect bugs at a much higher rate than before, and he believes it is essential to provide these capabilities to everyone as soon as possible to strengthen the ecosystem, as some malicious actors may attempt to exploit their own technology. Addressing recent incidents involving rogue AI agents, he acknowledges the risks but warns against an overreaction, considering security a challenging issue that can be resolved.

What is Released and How to Access It

Mistral has made the weights of Mistral Large 4 public, an open-source model with one trillion parameters. With this release, businesses and government agencies can use ML4 for their own purposes. The weights are accessible on Hugging Face and via an API. ML4 was first introduced earlier this month and carries the nickname "The Chonk."

Claimed Performance and Competition with Closed Models

Mistral claims that ML4 competes in performance with GLM-5.3, a benchmark for open models related to cybersecurity, while using 47% fewer tokens per task. The company also asserts that ML4 outperforms all other open models on agentic coding tasks according to the Artificial Analysis’ Coding Agent Index. This release comes amid increased competition between open-weight projects and closed offerings like those from OpenAI and Anthropic. Mistral advocates for an approach of independence for businesses and governments from external suppliers, deemed particularly relevant in cybersecurity, and Arthur Mensch summarizes this vision by presenting AI as a tool for empowerment. The publisher further claims to achieve top performance with a fraction of the computational resources and capital invested by closed labs.