Xiaomi unveils MiMo-V2.6 model after six-month focus on reinforcement learning

On September 17, 2026, Xiaomi's MiMo large model lead, Luo Fuli, announced the development of the MiMo-V2.6 model. Following the open-sourcing of MiMo-v2.5 in April 2026, Xiaomi remained inactive for nearly six months, dedicating time to exploring the extent to which reinforcement learning could be expanded. The model's training process involved approximately 20 billion tokens, with 1,568 prompts multiplied by 16 rollouts, using fully asynchronous methods. Luo also shared a live training link for MiMo-V2.6, revealing that the project has already cost over $1.25 million, equivalent to approximately 8.4 million yuan. The live stream indicates that the MiMo-V2.6 model is set for release.