Xiaomi Just Showed Us What Million-Dollar RL Training Looks Like

0:00 · Watch on YouTube

Xiaomi published MiMo-V2.6 Pro and Flash as a live, costed agent-reinforcement-learning run: a conventional trillion-parameter model was improved in about 30 RL steps for roughly $2.62 million (Pro) and $854,000 (Flas...

Read the full summary on retube