The Chonk goes on the offensive: Mistral open-sources the weights of its giant model

Le Chonk, Mistral AI's giant model, is making progress in agentic coding and cybersecurity. Its weights are now available for download.

Agentic programming and cybersecurity constitute the two main areas of advancement in the new version of Mistral Large 4, nicknamed "Le Chonk". The French laboratory is now making its weights available to developers and enterprises, who can adapt the model and run it on their own infrastructures. This evolution strengthens its value proposition in a market where technical performance is no longer the sole criteria for differentiation. Measurable Progress in Programming

With its 1 trillion parameters, including 52 billion active during inference when counting input and output layers, Mistral Large 4 retains its Mixture-of-Experts architecture. This new version advances primarily in its ability to work on complete software projects by analyzing code, proposing modifications, and verifying their execution.

The results published in Mistral AI's technical announcement report a first-place ranking among open-weight models on FrontierCode 2.0, Cognition's benchmark dedicated to evaluating programming agents. On the Artificial Analysis Coding Agent Index, the model reaches 48%, a result comparable to that of GLM-5.3, with an average token consumption 47% lower per task, according to the reported measurements. On Terminal-Bench 4.0, its score increases by 13 points to reach 41%. These performances remain specific to the protocols used and do not constitute a general ranking of all available models.

Three levels of reasoning—Low, Medium, and High—also allow users to adjust the resources deployed based on the difficulty of the requests, with the High level enabled by default. Cybersecurity Becomes a Competitive Ground

The other evolution involves computer security. The model reaches 51% on the Artificial Analysis Cyber Index, which specifically evaluates the identification, reproduction, and remediation of software vulnerabilities.

Several industry specialists, including CrowdStrike, Cloudflare, Rubrik, and Zscaler, are participating in evaluations or collaborations surrounding the model. The work focuses on code analysis, flaw detection, and assisting teams responsible for remediation.

However, the issue of malicious uses remains present. Mistral reports an average refusal rate of 96.6% against the offensive cyber requests in its tests, while acknowledging that the model's capabilities could facilitate certain real-world attacks. Consequently, a specialized variant for advanced security operations is subject to controlled access, reserved for vetted partners and authorized environments. An Open Model, but with Conditions

Beyond performance, access to the weights constitutes an important dimension of this new version. Available on Hugging Face, they allow organizations to host, adapt, and run Mistral Large 4 within their own environments. The Mistral Forge platform can also be used to customize the model based on specific data and needs.

This freedom, however, includes a commercial restriction. The Mistral Ecosystem Open-Weight license permits the use, modification, and commercialization of the model, but mandates a separate license for companies wishing to offer it commercially as a hosted service.

Developed and trained on European infrastructure, Le Chonk aligns with the sovereignty strategy championed by Mistral. Its new version combines technical performance, customization, and control over hosting—three dimensions that have become decisive for organizations wishing to maintain control over their AI systems.