Xiaomi
Xiaomi open-sources MiMo-V2.6 together with the RL stack used to train it
Xiaomi has released the Pro and Flash models of its MiMo-V2.6 series, a smaller 9B model, and the reinforcement learning stack it used to train them: more than 7,000 training environments, training code and a grading system. The company also disclosed what its final training phase cost.