Mistral previews Mistral Large 4, a trillion-parameter open-weight model with weights due by the end of October
The French company says the multimodal model beats every open-weight model from the US or Europe. Until the release, it is being red-teamed with security partners and state authorities.
Mistral launched a public preview of Mistral Large 4 on October 6 and says it will release the weights by the end of the month. The company calls the model its largest and most capable to date. It is a natively multimodal model with 1 trillion parameters, 49 billion of them active, and the preview API is available on Mistral Studio. Mistral says the model performs strongly on coding, agentic workflows and multimodal understanding. It claims the model is competitive with the strongest open models globally and significantly outperforms any open-weight model developed in the US or Europe. It also calls the model state of the art among open models on enterprise workloads in cybersecurity, finance and law, and says that in visual grounding it surpasses even frontier closed models. These are Mistral's own claims. Before the weights are released, the company is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities. A Japanese newsletter on note noted that US startup Reflection AI released its first open model a day earlier. It said both companies are targeting an open-model market where Chinese developers are strong. If the weights ship as planned, European and other users would have a top-tier option that is neither American nor Chinese and can run on their own servers.