Xiaomi Open-Sources MiMo-V2.6 Series, Ranking Highest Among Open-Weights AI Models
Xiaomi has officially released and open-sourced its MiMo-V2.6 native multimodal AI series, featuring Pro and Flash variants. The MiMo-V2.6-Pro model scored 46 on the Artificial Analysis Intelligence Index v4.3.2, surpassing Kimi K3 and GLM-5.3 to become the highest-ranked open-weights model. By scaling up reinforcement learning compute and releasing the underlying RL environments and frameworks, Xiaomi demonstrates how open-source models are rapidly closing the gap with top-tier proprietary LLMs. It also dramatically lowers the cost of state-of-the-art multimodal agentic capabilities for developers worldwide. Training MiMo-V2.6-Pro cost roughly $2.62 million over less than 6 days, utilizing a 1-million-token context window and multi-task RL across coding, visual, cyber, and general domains. Alongside open weights, Xiaomi released over 7,000 RL task environments, an end-to-end RL training framework, and the MiMo-V2.6-Distill-Qwen-9B model.
## BACKGROUND
The Artificial Analysis Intelligence Index is a composite benchmark aggregating evaluations across mathematics, science, coding, and reasoning to assess AI model capabilities. Recursive self-improvement (RSI) refers to AI systems improving their performance over successive iterations using automated feedback loops and reinforcement learning techniques.