MiMo V2.5 Pro

MiMo V2.5 Pro is Xiaomi's flagship text reasoning model optimized for the agent era, built on 1.02 trillion parameters with 42 billion active per token. Released in April 2026, it brings notable gains over its predecessor in general agentic capability, complex software engineering, and long horizon task completion. The model sustains long trajectories spanning hundreds to thousands of tool calls while holding coherence across its full 1 million token context window. Training used a three phase post-training recipe: supervised fine-tuning on curated data, domain specialized reinforcement learning, and multi-teacher on-policy distillation that merges specialist strengths across coding, reasoning, and tool use. Its hybrid attention architecture interleaves sliding window and global attention at a 6:1 ratio, with three lightweight Multi-Token Prediction layers that roughly triple output throughput at inference. MiMo V2.5 Pro reaches frontier level performance while spending roughly 40 to 60 percent fewer tokens than comparable models on benchmark trajectories, shown through tasks like building a Rust compiler and a full featured desktop video editor running to thousands of lines of code. It is open-sourced under the MIT license on Hugging Face, allowing commercial deployment, continued training, and unrestricted fine-tuning.