量子位
1 sourcesXiaomi Livestreams MiMo-V2.6 Reinforcement Learning Training, Burning $30K per Hour
Xiaomi publicly livestreamed the reinforcement learning training of its MiMo-V2.6 Flash and Pro models, showing real-time training costs, reward curves, and hardware status. After six months of silence, lead Luo Fuli explained that the team focused on how far reinforcement learning can scale, extending RL along three dimensions: training compute, environments/harnesses, and grader compute. After about a day of training, both models have shown clear capability improvements during RL.