Xiaomi is live-streaming an AI training run

Go to mimo.xiaomi.com/rl right now. You can watch a frontier AI model being trained, step by step, with live reward scores, task composition breakdowns, and running compute cost — all public.

Xiaomi's lead researcher Fuli Luo (ex-DeepSeek) surfaced after six months of silence to say the team spent that time on one question: how far can reinforcement learning scale? MiMo-V2.6 is their answer in progress. Each training step fires 1,568 prompts × 16 rollouts — roughly 2 billion tokens, fully async. One external analyst put the burn rate at ~$493k/day for the Pro run, $247k/day for Flash.

Neither OpenAI nor Anthropic publishes anything close to this. A phone company just set a new transparency bar for frontier AI training.