Today, we’re launching a public preview of Mistral Large 4. Unofficially ML4, very officially: le Chonk. ML4 pushes the frontier of open-weight performance. Preview API available on Mistral Studio. Weights drop end of this month.
ML4 is a 1 trillion-parameter natively multimodal model with 49 billion active parameters. It is Mistral’s largest and most capable model to date.
The model demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding. It achieves performance competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe. On critical enterprise workloads, including cybersecurity, finance and law, it is state-of-the-art among open models.
ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe. The public preview is served on that same infrastructure.
Cybersecurity
On the Artificial Analysis Cyber Index, ML4 ranks among the top five models globally and leads open-weight models developed outside China. On one test reproducing real vulnerabilities in open-source software, ML4 scores 82%—highest of any model. It solves 93% of Cybench challenges. Several leading closed models including Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse to perform the task.
Agentic Coding
ML4 scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Combined Coding Agent Index score of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. Blind human evaluation with Surge AI ranked ML4 Preview second of five models (3.74), behind only Claude Opus 5 (4.22).
Agentic Workflows
On AutomationBench (657 business workflows across Gmail, Google Sheets, Slack, Salesforce), ML4 scores 59.9%. On AA-Briefcase for long-horizon knowledge work, it reaches 1,393 Elo.
Multimodal
Strong visual grounding—surpassing GPT-6-Astra on Dense 200 (42% vs 41%). Handles complex documents, charts, satellite imagery, and engineering drawings.
Knowledge Work
Third-party evaluators (vals.ai) found ML4 exceeds GPT-6-Astra on legal and financial benchmarks. On HarveyAI’s Legal Agent benchmark, ML4 outperforms all open-source models.
Safety
On Lakera’s B3 AI Security Benchmark, ML4 resists 93.3% of attacks. KORA Benchmark score 1.691 among highest for OSS models.
Infrastructure
Funded by Mistral’s €3 billion Series D. Reinforcement learning run produces roughly 33 billion tokens per day on 3k GPUs. Weights release planned end of October 2026 with architecture details and benchmarks.