Xiaomi live-streams MiMo-V2.6 reinforcement learning training

Xiaomi is providing unprecedented transparency into its artificial intelligence development by live-streaming the reinforcement learning training process of its MiMo-V2.6 model. Through a public dashboard, the company is displaying real-time data including reward curves, GPU failures, context length, and infrastructure events.

The live stream also includes a running cost counter, revealing the massive economic scale of large-scale AI training. Reports indicate the training costs approximately $30,000 per hour. After 36 hours, costs were recorded at roughly $940,000, with later updates showing total expenditures exceeding $1.25 million.

Luo Fuli, head of the Xiaomi MiMo team and a former core author of DeepSeek V2, stated that the team has spent the last six months focusing on the scalability of reinforcement learning. The MiMo-V2.6 training process involves processing approximately two billion tokens per step across 1,568 tasks. The team plans to open-source technical details regarding their advancements in computation, environment/tool integration, and grader compute in the coming weeks.