Le Chonk: Built for Coders, Trusted by Defenders.

Mistral AI has released Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, following a public preview that drew strong developer and customer engagement. The model has since undergone reinforcement learning improvements, significantly enhancing its coding and defensive cybersecurity capabilities. These gains have accelerated without signs of saturation, prompting the release of its updated weights today. The improvements include substantial reasoning efficiency gains, reducing inference costs while maintaining high performance.

Mistral Large 4 leads open-weight benchmarks in coding, ranking first on FrontierCode and Terminal-Bench, and ties for top performance on the Artificial Analysis Coding Agent Index. It achieves a 19.7% score on Cognition's FrontierCode 2.0 benchmark, assessing end-to-end code quality and repository maintainer acceptance. The model also reduces token usage by 47% compared to leading open-weight models, making it more economical for large-scale coding agents.

In cybersecurity, Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index, demonstrating its ability to discover, reproduce, and patch vulnerabilities without breaking existing functionality. It maintains a 96.6% refusal rate on malicious cyber requests, the highest among open models, while retaining strong defensive security performance. The model has been validated by Booz Allen in source code security analysis, matching vulnerability detection recall rates of other frontier cyber models.

Mistral Large 4 supports three reasoning modes to balance performance and cost, and is fully open-weight with European training infrastructure. Organizations can deploy it sovereignly, customize it, and audit its architecture. ENISA is testing the model to identify vulnerabilities, audit code, and analyze defensive logs, highlighting its role in advancing cybersecurity through transparency and control.