AI news ·
Mistral unveils large 4 AI model with open-weight release planned for October
Mistral released a preview of its largest model, trained on 3,800 NVIDIA Grace Blackwell GPUs. The open-weight model, set for release by October's end, scores 93.3% on a key security benchmark.
Mistral has released a preview of its largest model yet, trained on 3,800 NVIDIA Grace Blackwell GPUs in European data centers. The model, officially named Mistral Large 4 and referred to as "le Chonk," handles instruction, reasoning, and agentic workflows in a single system and will have its weights released by the end of October after red-team testing by cybersecurity experts and state authorities.
Training and multilingual reach
The model was trained on multilingual data spanning more than 160 languages, including all official EU languages. Mistral emphasizes the open-weight and self-deployment capabilities, positioning the system as a sovereign AI option for industries where provider-level refusals could block critical work. Cybersecurity research is one example the company points to, where a model's willingness to comply with legitimate vulnerability probes matters.
Mistral's training approach uses what it calls Reinforcement Learning at scale, a methodology that adapts to the model's changing capabilities during training. This supports complex, long-horizon tasks across multiple domains rather than narrow, single-session benchmarks.
Performance across technical and professional benchmarks
Mistral Large 4 ranks in the top five globally on the Artificial Analysis Cyber Index. On coding tasks, it leads open-weight models on DeepSWE v1.1 and SWE-Atlas-QnA. For multimodal reasoning, the model scores above GPT-6 Astra on Dense 200, covering visual grounding, satellite imagery analysis, and technical drawing interpretation. It also outperforms GPT-6 Astra on HarveyAI's Legal Agent benchmark for financial analysis and achieves state-of-the-art results on SciCode-Verified for scientific workflows.
Security testing shows a 93.3% success rate on the Lakera B3 AI Security Benchmark, with strong resistance to prompt injections and refusal of malicious cybersecurity requests. Mistral is conducting red-team testing with cybersecurity leaders, vetted partners, and state authorities before the full deployment.
Access and upcoming release
A preview API is available now through Mistral Studio. The weights are scheduled for release by the end of October. Professionals working in fields that demand deep code analysis, financial reasoning, or multimodal document review can test the model immediately, with the option to self-deploy once the open weights drop.
Why this matters for IT and development professionals
For cybersecurity teams, a model that refuses malicious requests while allowing legitimate vulnerability research removes a friction point that some closed-provider APIs create. Developers working on agentic systems get a single model that handles reasoning, coding, and tool use without switching between specialized systems. The October weight release means teams can plan on-premise deployment for sensitive workflows where data cannot leave their infrastructure. The model's performance on SWE-benchmarks and the Lakera security test gives a concrete baseline to evaluate against current tools before committing to integration.