Beating AI News Flash: Both reinforcement learning training runs publicly livestreamed by Xiaomi MiMo have now been halted. The last fully completed runs for both Pro and Flash were 30 training steps each, with the two runs costing a combined $3.4747 million. Pro cost $2.6207 million, while Flash cost $854,000.The gains bought with money are obvious. On DeepSWE v1.1, Pro rose from 58.41 at step 1 to 72.57, an increase of 14.16 points; Flash rose from 48.67 to 65.68, an increase of 17.01 points. But both curves showed clear pullbacks in the middle, and it was not the case that the longer the training, the higher the score climbed all the way.The other two dashboard evaluations also rose. Pro's internal code evaluation increased from 57.54 to 65.43, and AutomationBench rose from 45.2 to 53.1.The training process was not smooth either. Pro triggered GPU OOM due to uneven expert load, and the training cluster and scorer also disconnected and restarted; Flash reran from step 15 due to an infrastructure error. Later, Xiaomi also filtered out tasks that had already become too easy for Pro and removed the cyber dataset, which showed abnormal rollout patterns, from Pro's subsequent training.