Mistral Releases Open Weights for Trillion-Parameter Large 4 Targeting On-Prem Cybersecurity
Frontier AI models · Enterprise AI
French AI startup Mistral AI has released open weights for its 1-trillion-parameter flagship model, Mistral Large 4, designed for cybersecurity defense and sovereign deployment. Organizations can host the model on-premise to mitigate business continuity risks associated with cloud service access cuts. While early metrics show strong resistance to malicious requests and solid coding agent performance, gaps remain compared to top closed-source systems.
French AI startup Mistral AI has officially released the open weights for its 1-trillion-parameter flagship model, Mistral Large 4 (nicknamed Le Chonk), aimed at allowing enterprises and public sector organizations to run AI locally to safeguard access sovereignty and operational continuity. According to Mistral, this updated release can be hosted on private infrastructure using inference frameworks like vLLM and SGLang, alongside API access. Tailored for cyber defense, finance, and manufacturing, the model is built to aid organizations in finding and remediating security vulnerabilities. The company reported a 96.6% average rejection rate for malicious cybersecurity requests and a 48% score on the Artificial Analysis Coding Agent Index. Addressing risks that open weights could aid cyberattacks, Mistral acknowledged the dual-use concerns, stating it monitors telemetry to refine guardrails and restricts specialized security builds to vetted design partners. Benchmarks reported by Fortune show Large 4 achieving 41% on Terminal-Bench 4.0, still trailing top closed-source models such as Anthropic's Mythos 5.1 (60.9%), indicating that operational value will depend on real-world verification.
Key takeaways
- Mistral AI released weights for the 1-trillion-parameter Large 4, enabling on-premise execution to mitigate third-party supply disruption risks.
- Positioned for cyber defense, Mistral reported a 96.6% malicious request refusal rate and a 48% Coding Agent Index benchmark.
- To counter potential misuse of open weights, Mistral is adjusting guardrails while restricting specialized security builds to vetted partners.
- In Terminal-Bench 4.0 evaluations, Large 4 scored 41%, leaving a measurable performance gap compared to Anthropic's Mythos 5.1 at 60.9%.
Source：科技報橘 Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.