Luo Fuli held it in for half a year, and Xiaomi live-streamed MiMo-V2.6 reinforcement learning directly.

Beating AI News Flash: Xiaomi is publicly livestreaming the reinforcement learning training of its next-generation MiMo-V2.6.

Lead Luo Fuli stated that the team has been silent for nearly half a year, focusing on researching how far reinforcement learning can still scale. Currently, V2.6 is still in training, and the specific technical solutions will be open-sourced in the coming weeks.

This time, the scale was ramped up very high from the start. Each step uses 1,568 prompts, with each generating 16 rollouts, totaling approximately 2 billion tokens, all running asynchronously. Agent tasks such as coding, general, vision, cybersecurity, and chat are also mixed into the same training session, with multiple harnesses used simultaneously.

The public can see real-time internal metrics for each step, including task composition, context length, and training time. UC Berkeley PhD student Yichuan Wang observed from the dashboard that each step involves about 1,500 tasks, with coding tasks accounting for about two-thirds, and over 40,000 active sandboxes at the same time.