Xiaomi Previews MiMo-V2.6 AI Model via Live-Streamed Reinforcement Learning Training
Xiaomi's MiMo team announced the upcoming release of MiMo-V2.6 and launched a public live stream of its training run, which has already accumulated over $1.25 million USD in compute costs. Led by Luo Fuli, the team spent six months researching the scaling limits of reinforcement learning across compute, agent environments, and grader scoring. Publicly live-streaming an ongoing LLM training run and sharing technical insights into agentic RL scaling brings rare transparency to frontier AI development. If Xiaomi's approaches to multi-task agentic RL and grader compute succeed, they could offer effective open techniques for training reasoning agents. MiMo-V2.6 processes approximately 2 billion tokens per step using 1,568 prompts and 16 rollouts in a fully asynchronous architecture. It introduces Agentic In-group Credit Assignment and scales compute dedicated to the grading process to reward multi-step reasoning and tool usage.
## BACKGROUND
Reinforcement learning fine-tuning relies on automated reward models or graders to evaluate model outputs and guide training. Credit assignment in agentic RL is the challenging process of determining which specific step or tool call in a long multi-turn interaction contributed to the final success or failure.