Mistral releases Le Chonk with 1 trillion parameters for coding and security

A new frontier for open weights

Mistral AI has officially released the weights for Mistral Large 4, which the company has nicknamed Le Chonk. This 1-trillion-parameter model is designed to be a flagship for open-weight AI, providing a balance between high-level reasoning and practical utility. It is natively multimodal, meaning it can process different types of data inputs beyond just text.

The model was developed and trained from scratch using European infrastructure. This specific lineage is intended to provide organizations with a foundation for sovereign AI, where they can own and control the underlying technology. By releasing the weights, Mistral allows developers to host the model on their own hardware rather than relying solely on an external API.

Mistral Large 4 aims to unify instruction-following, reasoning, and agentic capabilities into a single architecture. It is fluent in over 160 languages and is positioned as a state-of-the-art option for industries like finance and manufacturing. The release follows a public preview period where the company gathered feedback to refine the model's performance.

Optimized for coding and agents

The model shows significant improvements in coding tasks, particularly for agentic workflows where an AI performs multi-step actions. On the FrontierCode 2.0 benchmark, which measures if a human maintainer would merge an agent's code change, Mistral Large 4 achieved a score of 19.7%. It also leads the Terminal-Bench 4.0, which tests terminal-based coding workflows.

Efficiency is a core focus of this release, as the model is designed to reduce the number of tokens used during complex tasks. On the Artificial Analysis Coding Agent Index, it tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens on average. This makes it more economical for companies looking to run sophisticated coding agents at a large scale.

To give users more control over the cost-performance tradeoff, the model supports three distinct reasoning modes. These modes allow developers to choose how much computational power to dedicate to a specific task. This flexibility helps in managing inference costs while still maintaining high-quality outputs for different types of software engineering needs.

Defensive security and safety

Mistral AI worked with cybersecurity vendors to test the model's ability to handle vulnerability reproduction and patching. It scored 51% on the Artificial Analysis Cyber Index, which evaluates a model's ability to find, validate, and fix flaws in a codebase. Importantly, the model is designed to perform these defensive tasks without making unnecessary refusals.

Safety guardrails were also updated to distinguish between legitimate defensive security work and malicious, offensive requests. The model achieved an average refusal rate of 96.6% on malicious cyber requests, which is the highest rate measured among open models. This ensures that the model remains a tool for defenders rather than a utility for attackers.

For specialized needs, Mistral is offering a variant of the model optimized for advanced workflows like penetration testing and red-teaming. This version is subject to rigorous vetting and is only available to enterprise design partners in controlled environments. Other customers can use Mistral Forge to tailor the model with their own proprietary security data and detection logic.

Sovereignty and deployment

A major theme of this release is technological sovereignty, which refers to an organization's ability to retain control over its AI models. Because Mistral Large 4 is an open-weight model, organizations can inspect, customize, and deploy it on their own private infrastructure. This removes the dependency on Mistral to serve the model through a public API.

For those who prefer managed services, the model is available through Microsoft Foundry on Azure. This partnership provides private deployment options for developers who need enterprise-grade scaling. Both Microsoft and Mistral AI stated they are committed to expanding customer choice for next-generation AI models.

The model is released under the Mistral Ecosystem Open-Weight license, which is a modified MIT license. It allows for free use, modification, and commercialization, provided that businesses offering a Model-as-a-Service do not use it without a separate commercial license. This framework is designed to support both individual developers and large-scale institutional users.