Mistral has launched a public preview of Mistral Large 4, its largest and most capable open-weight model to date, and introduced the unofficial nickname "le Chonk." The model is natively multimodal and currently available through the Mistral Studio API, with full weights scheduled for release at the end of the month. Mistral says the preview is already competitive with the strongest open-source systems globally and superior to any open-weight model developed in the United States or Europe.
Frontier Architecture and Deployment
ML4 is a 1 trillion-parameter model with 49 billion active parameters and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European data centers. The public preview is served on that same infrastructure, reinforcing the company's investment in sovereign AI capabilities. Training data spanned more than 160 languages, including every official language of the European Union.
Cybersecurity Performance
On the Artificial Analysis Cyber Index, ML4 ranks among the top five models globally and leads open-weight models developed outside China. It resolves 93% of Cybench challenges and scores 82% on a test requiring reproduction and patching of a real software vulnerability. Mistral notes that several closed models refuse this task, while ML4 is intended to support legitimate vulnerability research and incident response under customer control.
Coding and Agentic Workflows
In software engineering, ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8%. That places it ahead of DeepSeek V4 Pro and Qwen3.8 Max, while a blind Surge AI evaluation ranked it second of five models at 3.74, behind Claude Opus 5. For agentic work, ML4 attains 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase.
Multimodal, Science, and Knowledge Work
ML4's multimodal reasoning includes visual grounding that surpasses GPT-6-Astra on Dense 200, with scores of 42% versus 41%. The company highlights applications in engineering, manufacturing, and earth observation involving satellite imagery, technical drawings, and complex documents. In science, mathematics, and knowledge work, Mistral says third-party legal and financial evaluations place the model ahead of GPT-6-Astra, with state-of-the-art open-weight results on SciCode-Verified and HarveyAI's Legal Agent benchmark.
Safety and Reinforcement Learning
Mistral reports that ML4 resists 93.3% of attacks on Lakera's public B3 AI Security Benchmark and ranks at its highest measured score among open-source models on the KORA safety benchmark. The company also says its refusal rates for malicious cyber prompts exceed those of all open-source systems tested. The reinforcement learning infrastructure behind the model uses 3,000 GPUs and produces roughly 33 billion tokens per day, with around 16 billion trainable completion tokens after filtering.
European Sovereignty and Roadmap
Mistral describes ML4 as a milestone funded by its €3 billion Series D, the largest equity round ever raised by a European technology company. The model was trained and is served from the company's own European data centers, supporting deployment under European law. Weights will be released by the end of the month, with additional details on architecture, benchmarks, and post-training methodology expected at that time.
Mistral Large 4 positions the company as a serious challenger in open-weight AI by pairing frontier performance with European infrastructure and deployment autonomy. Its emphasis on cybersecurity, agentic coding, multimodal reasoning, and professional knowledge work addresses areas where regulated or security-sensitive organizations often require self-deployment. With the Series D capital now being deployed, Mistral expects rapid improvements and a new generation of specialized models built on this foundation.