MiMo-V2.5: 310B total / 15B active (Sparse MoE); MiMo-V2.5-Pro: 1.02T total / 42B active (MoE)
1M tokens (Base variants: 256K)
MIT
Official open-source MiMo-V2.5 family release covering both the native omnimodal MiMo-V2.5 track and the larger MiMo-V2.5-Pro agentic track, each published with Base and full post-trained variants. Xiaomi positions MiMo-V2.5 as a unified text-image-video-audio model built on the MiMo-V2-Flash backbone with dedicated in-house vision and audio encoders, 310B total parameters with 15B active, roughly 48T training tokens, and a 1M-token context window. The Hugging Face model card reports a 729M vision encoder, a 261M audio encoder, and three Multi-Token Prediction layers. Xiaomi positions MiMo-V2.5-Pro as the flagship long-horizon coding and agent model, with 1.02T total parameters, 42B active parameters, hybrid attention, three Multi-Token Prediction layers, 27T pretraining tokens, and a 1M-token context window. Both tracks use staged training with supervised fine-tuning, agentic reinforcement learning, and Multi-Teacher On-Policy Distillation. Release materials emphasize frontier-level open-source performance in agentic workflows, long-context coherence, coding, and multimodal understanding; Xiaomi cites Claw-Eval gains, strong token efficiency, and autonomous runs spanning hundreds to thousands of tool calls, including a SysY compiler task completed with a perfect hidden-test score and a full video editor generated end to end.