Overview
- Xiaomi’s MiMo‑V2.6 is in an intermediate reinforcement learning training phase and the team is publicly streaming the run with the livestream page signaling an imminent release.
- The team expanded core infrastructure to support larger RL workloads, reporting roughly 2 billion tokens per training step in a fully asynchronous setup with 1,568 prompts and 16 rollouts.
- MiMo‑V2.6 adds broader environment and tooling to run multi‑task and agentic RL experiments and includes a dedicated “Grader Compute” cluster that allocates extra compute to scoring and reward assignment.
- Xiaomi reported cumulative training spending has passed $1.25 million, which the report notes is about ¥8.4 million at current exchange rates.
- The company says it will gradually open‑source the RL training details and materials over the next few weeks, a move that could let outside researchers inspect the team’s techniques and reproduce the experiments; Xiaomi previously open‑sourced MiMo‑v2.5 in April 2026.