According to reports on October 30th, European artificial intelligence leader Mistral has officially opened its new flagship model, codenamed “Le Chonk” — Mistral Large4 — to the public. As one of the largest open-source large models currently available, it has up to 1 trillion parameters and was trained from scratch using European local infrastructure. This marks a significant breakthrough in open-source large models as they move towards trillion-scale capabilities and native multi-modal integration.
Performance Hardening and Token Efficiency: Rejecting mere score boosting, focusing on ultimate cost-effectiveness
In terms of core capabilities and benchmark tests, Mistral Large4 demonstrates extremely strong competitiveness. In the private benchmark FrontierCode2.0 by Cognition (which mainly evaluates end-to-end code quality and the maintainers' willingness to merge code libraries), the model achieved a high score of 19.7%, placing it among the top models in current open-source weighted models. In the Coding Agent Index evaluation by Artificial Analysis, it obtained a score of 48%, successfully matching the industry-leading GLM-5.3.
It is particularly notable for its excellent Token efficiency. While achieving the same top-tier performance, Mistral Large4 uses 47% fewer Tokens per programming task on average compared to competitors. In the Terminal-Bench4.0 test, its performance improved by 13 percentage points to 41%, and the number of output Tokens was also reduced by 30%. This ability to “do more with fewer Tokens” directly addresses the major cost issue in current enterprise AI usage, offering developers a highly cost-effective option.
Three Mode of Reasoning and Reinforcement Learning: Balancing Performance and Cost
To enable developers to freely balance performance and cost in different scenarios, Mistral conducted an additional three-week reinforcement learning (RL) training for the model. The official platform offers three different inference modes, corresponding to simple tasks, medium complexity, and the most challenging scenarios that require high-precision thorough reasoning. The development team said that after the extended reinforcement learning training, the overall benefits of the model are still accelerating, and no signs of performance saturation have been observed.
Security Defenses and Detailed Barriers
In terms of security and compliance governance, Mistral Large4 also delivered a solid performance. In the Artificial Analysis Cyber Index evaluation, its score reached 51%, which mainly assesses the model’s ability to identify vulnerabilities in the code library, reproduce and verify them, patch them without compromising existing functions. Meanwhile, in tests for jailbreak and harmful prompt words such as JailbreakBench, StrongREJECT, and AgentHarm, its average rejection rate was as high as 96.6%, further improving from the 95% of the preview version.
It is worth mentioning that the official side has shown a pragmatic approach in its security strategy: the model is firm in rejecting black-box attacks targeting online targets, and it also recognizes the necessity of using related technologies for legitimate defense and penetration testing. As a result, Mistral has provided preview access to specific variant versions designed for enterprises in the cybersecurity field such as Orange Cyberdefense, but all operations must be subject to strict contractual constraints and independent or authorized environment reviews.
Ecosystem Layout and Open Source License Agreement
In terms of application ecosystem support, Mistral Large4 has received backing from many giants in the fields of cybersecurity and sovereign cloud, such as Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, and NVIDIA. Users can download model weights through Hugging Face, and deploy them locally using vLLM or SGLang; additionally, it has been available on mainstream platforms like Microsoft Foundry (the first cloud provider to launch it), NVIDIA NIM containers, and OpenRouter. The officially published API pricing is $0.70 per million Token for input, $0.07 for cache input, and $2.10 for output.
Regarding the highly regarded open-source license agreement, this model adopts a modified version of the MIT agreement — the Mistral Ecosystem Open-Weight license. The agreement allows users to use, modify, fine-tune, and commercialize it freely, but it sets a core restriction: any enterprise that primarily operates as a “Model as a Service” (MaaS) must apply separately to Mistral for commercial permission before launching commercial use. This strategy not only maintains the vitality of the open-source ecosystem but also effectively protects its commercial advantages.