Xiaomi MiMo-V2.6 AI model development and release

Xiaomi initiated a transparent development process for its MiMo-V2.6 reinforcement learning model by live-streaming its training via a public dashboard. This stream provided real-time data on GPU failures, infrastructure events, and training costs, which were reported to exceed $1.25 million during the initial stages.

Following the training phase, Xiaomi officially released and open-sourced the MiMo-V2.6 series, which includes the Pro and Flash multimodal models. The release focused on scaling reinforcement learning compute to explore Recursive Self-Improvement. The Pro model achieved a score of 46 on the Artificial Analysis Intelligence Index, outperforming certain open-weight models while trailing top closed-source systems.

The new models feature capabilities in 3D spatial reasoning, multimodal perception, and agentic computer use, such as assisting in 3D modeling and robotic arm control. Xiaomi has open-sourced the model weights, technical reports, and over 7,000 high-quality reinforcement learning task environments.

Updated details regarding the release indicate that the MiMo-V2.6-Pro has tied with Grok 4.7 as the top-ranked open-weight model globally on the Artificial Analysis Intelligence Index. The models feature a 1-million-token context window and can process text, images, audio, and video. Xiaomi disclosed that the reinforcement learning training phase for the Pro model cost approximately $2.62 million, while the Flash model cost approximately $850,000.

However, the release has faced scrutiny; Anthropic has raised concerns, alleging that Xiaomi “misused Claude data for large-scale data collection during development.”