Mistral 开放 Large 4 权重，主打编码与网络安全

Mistral has released the weights of Mistral Large 4, codenamed Le Chonk, making its 1-trillion-parameter, natively multimodal flagship available as an open-weight model for coding, cybersecurity and enterprise agents. The public preview arrived earlier this month. Mistral also reports that three weeks of additional reinforcement learning drove further gains, including more efficient reasoning and lower inference costs.

The model scored 19.7% on Cognition’s private FrontierCode 2.0 benchmark. On Artificial Analysis’s Coding Agent Index, it reached 48%, matching GLM-5.3 while using 47% fewer tokens per task on average. Its Terminal-Bench 4.0 score rose 13 percentage points from preview to 41%, while output tokens fell 30%. Low, Medium and High reasoning modes now let developers tune cost against performance, with High used by default. In Surge Code’s blind comparisons, it ranked ahead of GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and behind only Claude Opus 5.5 among the models tested.

Cybersecurity is the second focus. Mistral Large 4 scored 51% on the Artificial Analysis Cyber Index, which covers finding, reproducing, validating and patching vulnerabilities without breaking existing functions. Mistral says its guardrails preserve defensive work while refusing an average of 96.6% of malicious cyber requests, the highest rate it measured among open models. The company also acknowledged potential uplift for real attacks in live-traffic monitoring and said it is modifying guardrails. A specialized cyber variant is being tested with vetted enterprise partners on owned or authorized systems.

Beyond coding and security, the model reached 73.3% on Finch for finance workflows, 84.5% on BrowseComp, up from 75% at the SFT checkpoint, and 97.1% attack resistance on the B3 Agent Security Benchmark, compared with 93.3% in preview. It is a hybrid instruct-and-reasoning mixture-of-experts model with multimodal input and native fluency in more than 160 languages.

Developed and trained from scratch on European infrastructure, Mistral Large 4 anchors Mistral’s push for sovereign AI: organizations can inspect, customize and run it on their own systems, or tailor it through Mistral Forge and build governed agents in Mistral Studio. It is available through Mistral Studio, the Mistral API, OpenRouter, Hugging Face, Microsoft Foundry and NVIDIA-hosted endpoints, with self-hosting through vLLM or SGLang.

The Mistral Ecosystem Open-Weight license is a modified MIT license that permits free use, modification, fine-tuning and commercialization. Model-as-a-Service businesses need a separate commercial license. API pricing starts at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Sources and related context

• Mistral Large 4 model documentation: Supports: Mistral's documentation details the model architecture, reasoning modes, context window, and API pricing described in the article.
• Mistral Large 4 open weights and model card: Supports: Mistral's model card provides the downloadable weights, deployment guidance, and open-weight license terms behind the release.

Source