Xiaomi has publicly halted its MiMo reinforcement learning project, with two training runs costing $3.47 million.

Beating AI News Brief: Xiaomi’s MiMo has halted two publicly live-streamed reinforcement learning training runs. Both the Pro and Flash runs completed a full 30 training steps each, with total combined costs of $3.4747 million. The Pro run cost $2.6207 million, while the Flash run cost $854,000. The investment delivered clear performance gains: On DeepSWE v1.1, Pro’s score rose from 58.41 at step 1 to 72.57, a 14.16-point increase; Flash’s score jumped from 48.67 to 65.68, a 17.01-point gain. However, both curves showed noticeable dips in the middle, meaning scores do not rise steadily with longer training. Two additional benchmark tests also improved: Pro’s internal code evaluation score rose from 57.54 to 65.43, and its AutomationBench score increased from 45.2 to 53.1. The training process was not smooth: Pro encountered GPU out-of-memory (OOM) errors due to uneven expert load, plus disconnections and restarts between its training cluster and evaluator. Flash had to restart from step 15 due to infrastructure errors. Later, Xiaomi filtered out tasks that were too simple for Pro and removed the cyber dataset with abnormal rollout patterns from Pro’s subsequent training.