Xiaomi broadcasts the real-time reinforcement learning process of its MiMo-V2.6 models via a public dashboard. The Pro version uses 1 trillion total parameters with 42 billion active, while the Flash model uses 310 billion total parameters and 15 billion active. Daily costs for the Pro run are $493,000, with each training step costing $70,000, while the Flash model costs $247,000 per day. Total expenditures for the project have exceeded $1 million.

The training scales to 2 billion tokens per step across 1,568 prompts and 16 rollouts. A SemiAnalysis review of the livestream metrics identifies a group size of 16, suggesting a GRPO algorithm variant, and an average rollout length of 93,000 tokens. The data comprises 26 datasets across code, visual, general, chat, agentic, and cyber categories. Xiaomi will open-source the technical details over the coming weeks.