Mistral Opens Large 4 Weights for Coding and Defensive Security

Introduction

Mistral has released the weights of an improved version of Mistral Large 4, also known as Le Chonk. The model entered public preview earlier this month, and the newly released version reflects roughly three additional weeks of reinforcement learning. Mistral describes it as a native multimodal model with about one trillion parameters, aimed particularly at agentic coding, complex reasoning, and defensive cybersecurity.

The announcement is notable less for parameter count alone than for its focus on usable efficiency. Mistral says the updated model improves its coding results while requiring fewer output tokens for comparable tasks. That matters for coding agents, which may call a model repeatedly, execute commands, inspect test failures, and revise a patch before a job is complete.

Key points

• Stronger repository-level coding results. Mistral reports a 19.7% score on Cognition’s private FrontierCode 2.0 benchmark. The test evaluates not only whether code is produced, but also whether the resulting change is good enough for a repository maintainer to merge. On Terminal-Bench 4.0, Mistral reports a score of 41%, a 13-point improvement over the public preview and a leading result among open-weight models.
• Efficiency is part of the pitch. On the Artificial Analysis Coding Agent Index, Large 4 is reported at 48%, effectively matching the leading open-weight model in the comparison while using 47% fewer tokens per task on average. Mistral also reports a 30% reduction in output tokens on Terminal-Bench 4.0. These figures could be important for teams running coding agents at scale, where repeated reasoning and long outputs can dominate serving costs.
• Three reasoning settings. Low is intended for simple tasks and replaces the former None mode. Medium is positioned for everyday work, while High is the default for difficult tasks requiring more precision and deliberate reasoning. The settings give developers an explicit way to allocate reasoning effort instead of applying the most expensive mode to every request.
• A defensive-security emphasis. Mistral reports a 51% score on the Artificial Analysis Cyber Index, which covers vulnerability discovery, reproduction and validation, and patching without breaking existing behavior. The company also says the model refuses an average of 96.6% of malicious cyber requests across the cited safety evaluations. Its stated goal is to preserve useful defensive capabilities while maintaining a stronger boundary around offensive misuse.
• Designed for controlled deployment. Mistral says the model was trained on European infrastructure and presents the release as a foundation for sovereign AI. Open weights can give enterprises, governments, and researchers more control over hosting, customization, data residency, and model inspection than a service that is available only through an external API.

Why it matters

Large 4 illustrates how competition among open models is moving beyond raw scale. For agentic coding, task economics depend on more than benchmark accuracy: token volume, the number of tool calls, and the model’s ability to recover from failed commands all affect the final cost. By exposing several reasoning modes, Mistral is treating reasoning budget as an operational control that can be tuned per request.

Cybersecurity introduces a more difficult balance. Defenders need models that can analyze vulnerabilities, reproduce issues, and draft repairs, yet the same capabilities can be redirected toward attacks. Mistral’s published Cyber Index score and refusal rate show an attempt to optimize both usefulness and safety boundaries. They should still be read as vendor-reported evidence rather than a complete independent assessment. The benchmark prompts, system configuration, hardware, and deployment context can all influence the outcome.

The practical significance of the release will therefore depend on more than the headline parameter count. Its combination of open weights, native multimodality, adjustable reasoning intensity, and coding and security specialization could make Large 4 attractive as a self-hosted foundation. Independent testing on diverse codebases, hardware configurations, and sustained security workflows will determine how well those advantages generalize.

Source: Hacker News