Mistral AI releases Large 4 open-weight model for cybersecurity

A private model for security

French startup Mistral AI recently launched its Large 4 model, which the company has nicknamed "Le Chonk." This new release is specifically designed to assist organizations in finding and fixing cybersecurity vulnerabilities within their systems. By focusing on security, Mistral aims to provide a tool for identifying flaws and performing agentic coding tasks.

What makes this release notable is the choice of an open-weight format. Unlike closed models like Google’s Gemini, which are primarily accessed through online services, ML4 can be downloaded and run independently. This allows companies to keep their data entirely on their own infrastructure, avoiding the risks of routing sensitive information through external servers.

As reported by Crypto Briefing, the model was trained using 3,800 Nvidia Grace Blackwell GPUs situated in data centers across Europe. This focus on local hosting is part of a broader push for digital sovereignty, ensuring that organizations are not dependent on a third-party provider that could change terms or revoke access at any time.

Technical scale and efficiency

ML4 is a massive system featuring roughly 1 trillion total parameters. However, the model uses a technique where only 52 billion parameters are active for any given token. This means that while the model is large, most of its parameters sit idle during each step of processing, which helps manage the computational load.

Despite its size, the model has a surprisingly modest hardware footprint. Mistral states that ML4 can operate on a minimum of four Nvidia B200 or B300 GPUs. This lowers the barrier for enterprises because it means they do not need a hyperscaler’s worth of hardware to run a trillion-parameter system in-house.

The model also features a 1 million-token context window. A context window is the amount of text a model can consider at one time. This high capacity allows the model to take in very large document sets or entire codebases in a single pass, which is a critical requirement for hunting for complex flaws across large software projects.

Benchmarks and performance

Mistral has positioned ML4 to compete with other leading models on security-specific benchmarks. On the CyberGym-E2E test, the model scored 79–82%. It also topped the best open-weight scores on Cybench, which is another suite used to test security capabilities. These results indicate that the model is highly capable of handling adversarial tasks.

The model showed significant improvement between its initial preview in early October and its full release. On the Terminal-Bench 4.0, its score jumped from 28% to 41%. Mistral attributes this growth to reinforcement learning that was shaped by direct feedback from corporate and government users during the development phase.

Additionally, ML4 has been shown to match or beat some leading Chinese models, such as GLM-5.3, on cybersecurity tasks while using fewer tokens. Mistral also emphasizes that the model has high refusal rates for malicious prompts. The goal is to create a tool that remains useful for defenders without becoming an assistant for attackers.

Adoption and accessibility

The weights for ML4 were made live on Hugging Face under a modified MIT-style license. This allows for broad deployment, but Mistral also provides an API route for those who prefer not to host the model themselves. The pricing for this route is roughly $0.70 for every million input tokens and $2.10 for every million output tokens.

This dual approach caters to different types of users. Smaller teams that may not have the budget for high-end GPUs can use the API to access the model's capabilities. Meanwhile, larger organizations with strict data-handling rules can choose to deploy the model on their own private infrastructure.

The release also provides a European alternative for enterprises concerned about the origin of open-weight models. By offering a model that is competitive with Chinese alternatives like GLM-5.3, Mistral provides a different path for companies seeking high-performance security tools. This move highlights the growing importance of regional AI development in the global market.