Introducing Mistral Large 4 | Mistral
Short summary
Mistral released a public preview of ML4 ("le Chonk"), a 1 trillion-parameter multimodal model trained on European infrastructure, claiming state-of-the-art performance on cybersecurity, coding, and knowledge work tasks compared to both open-source and closed models. Full model weights drop end of month after red-teaming with security partners.
- ML4 scores 82% on reproducing and patching real vulnerabilities (highest of any tested model) while closed models like Claude Opus 5.5 and GPT-6 Astra score near zero due to safety refusals, making it uniquely useful for defensive security work.
- The model was trained from scratch on Mistral's own European infrastructure (3,800 NVIDIA Grace Blackwell GPUs) and will run independently under European law, positioning it as a sovereignty play against US-controlled AI systems.
- On blind human evaluation of coding quality, ML4 ranked second (3.74/5) behind only Claude Opus 5 (4.22/5), and outperforms competitors on agentic workflow benchmarks (AutomationBench 59.9%) and knowledge work tasks like legal and financial analysis.
Full summary
Mistral just released a preview of Mistral Large 4 (ML4), a 1 trillion-parameter model with 52 billion active parameters that runs natively on images and text. They're calling it "le Chonk" internally. The weights drop at the end of the month, and you can test the preview API now on Mistral Studio. What makes this worth paying attention to: they trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in their own European datacenters, which matters for the sovereignty angle they're pushing hard. The model hits competitive performance with the strongest open-source models globally and outperforms any open-weight model from the US or Europe. In specific domains like visual grounding, it actually beats some closed frontier models.
The cybersecurity angle is where ML4 gets genuinely interesting. It scores 82% on the Artificial Analysis Cyber Index test that asks models to find and patch real vulnerabilities in actual software—the highest of any model tested. Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse to do it. Mistral's argument is straightforward: defenders need to prove flaws are real, which safety filters in closed models actively block. Since threat actors are jailbreaking those same models for offense, defenders need systems without those constraints. ML4 also handles malware analysis, vulnerability prioritization, and detection rule writing in ways it wasn't explicitly trained for. For organizations running security operations, they can deploy it on private infrastructure without depending on any provider's refusals or access restrictions.
Beyond security, the model shows strong performance across coding (61.7% on DeepSWE v1.1), agentic workflows (59.9% on AutomationBench), and multimodal tasks. In a blind human evaluation of code quality, professional annotators ranked ML4 second out of five models tested, behind only Claude Opus 5 (3.74 vs 4.22). It handles complex visual tasks like analyzing gigapixel satellite imagery and engineering drawings with zoom-and-verify capabilities. The training included data in over 160 languages, including every EU official language, and they worked with enterprises across finance, manufacturing, pharmaceuticals, and shipping to build it. More benchmarks and architectural details are coming as they finalize the weights.