MiMo-V2.6-Pro-MOPD
MiMo-V2.6-Pro-MOPDXiaomi’s flagship omnimodal MoE with 1.02T parameters, 42B active and 1M-token context, upgraded via MOPD for stronger agents and fewer repeated tool calls.
Model details
View repositoryMiMo-V2.6-Pro-MOPD is Xiaomi’s flagship omnimodal AI model, built on a sparse Mixture-of-Experts architecture with 1.02T total parameters (42B activated) and a hybrid sliding-window/global-attention design supporting contexts up to 1M tokens. It natively processes text, images, video, and audio through dedicated vision and audio encoders while producing text outputs. Unlike the RL checkpoint, which is trained through large-scale mixed reinforcement learning on verifiable tasks and agentic grading, the MOPD version further distills multiple domain-specialized RL and SFT teachers into the model on-policy. This extends its capabilities to harder-to-verify domains such as long-horizon game development, scientific research, and embodied intelligence, while reducing repetitive tool calls in agentic workflows.
