According to CoinMeta, Luo Fuli stated that Xiaomi is publicly live-streaming the reinforcement learning training for the next generation of MiMo-V2.6. The team has been silent for nearly half a year, focusing on researching the scalability of reinforcement learning. Currently, v2.6 is still in training, and the specific technical solutions will be gradually open-sourced in the coming weeks. Each step uses 1568 prompt, and each generates 16 rollout, totaling about 2 billion token. The entire process runs asynchronously. Tasks such as code, general purpose, vision, network security, and chat are combined in the same training session, and multiple sets of harness are used simultaneously. The outside world can see real-time internal indicators such as the composition of each task, context length, and training time. Doctoral student Yichuan Wang observed that there are about 1500 tasks per step, with code tasks accounting for about two-thirds, and more than 40,000 sandboxes are active at the same time.