Mistral releases open weights for Mistral Large 4 with strong cybersecurity benchmarks

Mistral AI has released the open weights for Mistral Large 4, its 1-trillion-parameter multimodal model also known as Le Chonk. The release follows a public preview and includes substantial post-training improvements in coding and defensive cybersecurity tasks.

Cybersecurity performance

The model scores 51% on the Artificial Analysis Cyber Index, which evaluates vulnerability discovery, reproduction, validation, and patching. In preview testing it reached 82% on a task that requires reproducing a real vulnerability and then patching it. It also achieves 41% on Terminal-Bench 4.0 after reinforcement learning gains, leading open-weight models on that benchmark while using fewer output tokens.

Mistral emphasizes that the model distinguishes defensive security work from offensive requests, with a 96.6% average refusal rate on malicious cyber prompts across standard benchmarks. A specialized variant with reduced moderation for vetted partners is available to organizations such as Orange Cyberdefense for penetration testing, vulnerability assessment, and red-teaming under controlled conditions.

European sovereignty focus

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers. Open weights are released under the Mistral Ecosystem Open-Weight License, allowing modification, fine-tuning, and commercial use with a restriction on operating Model-as-a-Service offerings without a separate commercial license from Mistral. European deployment options and self-hosting support are highlighted for organizations seeking control over sensitive security workflows.

Additional improvements include three reasoning modes for cost-performance tradeoffs, stronger agentic coding results, and gains on enterprise benchmarks such as financial spreadsheet tasks and web browsing. The weights are available on Hugging Face, with API access through Mistral Studio and other platforms.

Practical implications

Organizations that need auditable, self-deployed AI for security operations now have a high-performing open-weight option that avoids provider-level refusals common in closed models on legitimate defensive tasks. The release also underscores ongoing European efforts to build sovereign AI infrastructure independent of other cloud providers.

Sources

• mistral.ai
• mistral.ai
• www.orangecyberdefense.com

Cover photo by Pachon in Motion on Pexels, used under the Pexels License.