Xiaomi's MiMo-V2.6 Enters Mid-Training Phase, Live Stream Shows Over $1.25 Million Spent

Luo Fuli, head of Xiaomi's MiMo large model, said in a recent post that after open-sourcing MiMo-v2.5 in April, the team went quiet for nearly half a year to study how far reinforcement learning can scale. MiMo-V2.6 is now in the middle stage of reinforcement learning training.

The team expanded three capabilities for the model. For compute, each step uses about 2 billion tokens, with 1,568 prompts and 16 rollouts in a fully asynchronous setup. For environments and tools, it supports multi-task agent reinforcement learning that mixes multiple frameworks in one run. For grader compute, it adds more power to the scoring process itself, using Agentic In-group Credit Assignment with test cases and rubric rewards.

Luo also shared a live training stream for MiMo-V2.6, showing total spending has passed $1.25 million. The stream page says the model is coming soon. The team plans to open-source these details in the coming weeks.