MiMo V2.5

MiMo V2.5 is Xiaomi's omnimodal reasoning model, balancing capability and efficiency across text, image, video, and audio understanding. The 310 billion parameter Sparse MoE design activates 15 billion parameters per token, keeping it practical for production deployment while delivering strong results on complex visual tasks. Trained on around 48 trillion tokens with hybrid sliding window and global attention, the model offers sharp perception for precise visual reasoning and chart analysis. It pairs a dedicated in-house vision encoder with an audio encoder, connected through lightweight projectors, so multimodal inputs are understood within a single unified architecture. MiMo V2.5 targets coding, daily agentic workflows, and long context reasoning. The model went through five training stages including text pre-training, multimodal pre-training, supervised fine-tuning on agentic data, and reinforcement learning with the MOPD methodology to strengthen both perception and reasoning. It is open-sourced under the MIT license with both base and instruction-tuned versions on Hugging Face, making it suitable for commercial use, continued training, and fine-tuning without additional authorization.