Summary
Mistral has released a public preview of ML4 through its API and says it will publish the model’s weights on October 27. The French company positions ML4 as an open-weight European alternative to leading Chinese and closed American systems, with an emphasis on cybersecurity and defense. Mistral says the model supports more than 160 languages, was trained for two months on 4,000 Nvidia Grace Blackwell GPUs in European data centers, and performs strongly in multimodal work, coding, manufacturing, finance and grounding. Those performance claims have not yet been independently verified, and reinforcement learning was still being completed at launch. The company says developers, governments and security teams can test the API before the planned weight release.
Hana’s perspective
I find the promised release of the weights more important than the sovereignty framing. Animation and media teams can inspect, adapt and run an open-weight model under their own controls, which can help with confidential scripts, asset metadata and multilingual localization. But the preview arrives before the weights and before independent evaluation, while training is still being completed. That makes benchmark leadership a claim, not a settled result. I would judge ML4 after developers can test its real hardware requirements, licensing, multimodal behavior and safety controls in production—not by whether it is labeled European, Chinese or American.
Are you a writer with a great story? We'd love to hear from you: APPLY HERE