Introducing Mistral Large 4

Le Chonk is trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters, using multilingual data across 160+ languages, including all EU official languages. It is currently available via a preview API on Mistral Studio, with weights scheduled for release by the end of October. The model is being red-teamed by cybersecurity leaders, vetted partners, and state authorities before full deployment. Mistral emphasizes its open-weight and self-deployment capabilities, positioning it as a sovereign AI solution for industries like cybersecurity, where provider-level refusals could block critical vulnerability research. The company also highlights its Reinforcement Learning (RL) at scale methodology, which adapts to the model’s evolving capabilities and supports complex, long-horizon tasks across domains.

Mistral claims le Chonk is state-of-the-art in cybersecurity (ranking top five globally on the Artificial Analysis Cyber Index), coding (leading open-weight models on DeepSWE v1.1 and SWE-Atlas-QnA), and multimodal reasoning (surpassing GPT-6 Astra on Dense 200). It also excels in financial analysis (outperforming GPT-6 Astra on HarveyAI’s Legal Agent benchmark) and scientific workflows (state-of-the-art on SciCode-Verified). The model’s resistance to prompt injections and refusal of malicious cybersecurity requests are also highlighted, with a 93.3% success rate on the Lakera B3 AI Security Benchmark.