October 30 news: Europe's leading AI company Mistral has officially open-sourced its new flagship model codenamed "Le Chonk" — Mistral Large4. As one of the largest large models in the open-weight camp, it has up to 1 trillion parameters and was trained entirely from scratch on European local infrastructure, marking a substantive breakthrough for open-source large models on the path toward trillion-scale and native multimodal integration.
Head-on Performance and Token Efficiency: Rejecting Pure Benchmark Chasing, Focusing on Extreme Cost-Effectiveness
In terms of core capabilities and benchmarks, Mistral Large4 demonstrates extremely strong competitiveness. In Cognition's private benchmark FrontierCode2.0, which mainly evaluates end-to-end code quality and codebase maintainers' willingness to merge, the model scored a high 19.7%, ranking among the top current open-weight models. In Artificial Analysis's Coding Agent Index, it achieved 48%, successfully essentially matching the industry-leading GLM-5.3.
Particularly noteworthy is its outstanding token efficiency. While achieving equivalent top-tier performance, Mistral Large4 consumes on average 47% fewer tokens per programming task than competitors. In the Terminal-Bench4.0 test, its performance improved significantly by 13 percentage points over the preview version to 41%, while output token count also decreased by 30%. This "do more with fewer tokens" characteristic directly addresses the pain point of enterprise AI invocation costs, bringing developers a highly cost-effective choice.
Three-Tier Reasoning Modes and Reinforcement Learning: Balancing Performance and Cost
To allow developers to freely balance performance and expense across different scenarios, Mistral conducted an additional three weeks of reinforcement learning (RL) training for the model and officially provides three different reasoning modes, corresponding to simple tasks, medium complexity, and the hardest scenarios requiring high-precision, thorough reasoning. The development team said that after the extended reinforcement learning training, the model's overall gains are still accelerating, and no signs of performance saturation have been observed yet.
Security Defenses and Fine-Grained Guardrails
In security and compliance governance, Mistral Large4 has also delivered solid results. In Artificial Analysis Cyber Index, it scored 51%, mainly assessing the model's ability to discover vulnerabilities in codebases, reproduce and verify them, patch them without breaking existing functionality. Meanwhile, in jailbreak and harmful-request prompt tests such as JailbreakBench, StrongREJECT, and AgentHarm, its average refusal rate reached 96.6%, further improving from 95% in the preview version.
It is worth mentioning that the official security strategy shows a pragmatic attitude: the model is firm in refusing black-box attacks against online targets, while also recognizing the necessity of related technologies being used for legitimate defense and penetration testing. To this end, Mistral has opened preview access to specific variant versions specifically for enterprise design partners in cybersecurity such as Orange Cyberdefense, but all operations must undergo strict contractual constraints and review in owned or authorized environments.
Ecosystem Layout and Open-Source License
In terms of application ecosystem support, Mistral Large4 has received backing from many giants in cybersecurity and sovereign cloud, including Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, NVIDIA, and others. Users can download the model weights via Hugging Face and self-deploy locally via vLLM or SGLang; it has also landed on mainstream platforms such as Microsoft Foundry (the first cloud provider to list it), NVIDIA NIM containers, and OpenRouter. The officially announced API pricing is $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output.
Regarding the closely watched open-source license, the model uses a modified MIT license — the Mistral Ecosystem Open-Weight License. The license allows users to freely use, modify, fine-tune, and commercially deploy it, but sets one core restriction: any company primarily engaged in "Model-as-a-Service" (MaaS) must separately apply to Mistral for a commercial license before commercial use. This strategy maintains the vitality of the open-source ecosystem while also effectively protecting its own commercialization moat.
