Mistral Opens Large 4 Weights, Emphasizing Coding and Cybersecurity
Mistral has released open weights for Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, after public preview feedback drove improvements in coding and cybersecurity through additional reinforcement learning. The model leads open-weight coding benchmarks: 19.7% on FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index while consuming 47% fewer tokens than competitors, and 41% on Terminal-Bench 4.0 with 30% fewer output tokens. For defensive cybersecurity—detecting and patching vulnerabilities—it achieves 51% on the Artificial Analysis Cyber Index while rejecting malicious requests at 96.6%.
The release emphasizes “sovereign AI” architecture: European-trained with open weights enabling on-premises deployment and customization. Mistral grants select enterprise partners a specialized variant for controlled penetration testing and red-teaming. Performance extends beyond security and coding to web browsing (84.5%), financial workflows (73.3%), and agent security, resisting 97.1% of attacks. Through Mistral Forge fine-tuning, organizations can optimize task-specific variants. Adoption by mainstream institutions like La Banque Postale signals rising demand for independently-trained, open-source alternatives to proprietary closed systems.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.