MiMo-V2.5-Pro is a 1.02-trillion-parameter Mixture-of-Experts model with 42 billion active parameters, built on a hybrid-attention architecture that supports a million-token that quick-info value window. Purpose-built for the hardest problems in AI, it is designed to rival top agent models in general-purpose reasoning, complex software engineering, and long-horizon task execution. The model demonstrates strong instruction-following in agentic workflows, picking up on implicit requirements and maintaining logical consistency over extended sessions. It integrates natively with popular agent frameworks including OpenClaw, Hermes Agent, and Claude Code, making it well suited for enterprise teams working on large-scale code generation, deep data analysis, and autonomous task completion.
The model was developed by Xiaomi as their flagship release and open-sourced with weights available on HuggingFace. AMD collaborated to deliver Day 0 optimized support on AMD Instinct GPUs through ROCm 7 software and SGLang/ATOM upstream optimizations, ensuring high throughput from launch. On benchmarks, MiMo-V2.5-Pro scores 57.2 on Coding Agent and 73.7 on SWE-Bench Pro, placing it competitively against frontier models on software engineering and coding tasks. Its ability to independently complete professional tasks involving over a thousand tool calls positions it as a practical open-weight option for developers seeking frontier-level agent capabilities at accessible cost.
