Xiaomi trains MiMo 2.6 in public: RL run cost passes $1M

Xiaomi is streaming reinforcement-learning post-training of MiMo 2.6 on a public dashboard: MiMo 2.6 Pro (~1T parameters, 42B active) and 2.6 Flash (309B, 15B active). The Pro run's cost counter passed $1,048,236 at $5.71 per second, with seven restarts in 51 hours and pass rate rising from 56.5% to 61.5%.

• MiMo 2.6 Pro: ~1T parameters, 42B active, cost counter at $1,048,236
• Pro run burns $5.71 per second, about $20,500 per hour
• Pass rate rose from 56.5% to 61.5% over 14 steps
• Seven restarts in 51 hours, including removal of the cyber dataset