MiMo-V2.5-Pro

MiMo-V2.5-Pro is a sparse Mixture-of-Experts model designed for complex, long-running work. Its hybrid-attention design activates 42 billion of 1.02 trillion parameters, aiming to provide substantial capacity without activating the full model on every token. The official documentation also describes its reach as comparable to leading systems in high-intensity agent scenarios.

The model is best suited to agentic software engineering and other tasks that require sustained reasoning across extensive context. Xiaomi reports gains over its predecessor in general agent behavior, complex coding, and long-horizon completion. On the company’s shared Coding Agent comparison, it records 57.2, narrowly trailing two stronger listed results while exceeding the other models shown, making it a practical option for open-weight development of capable coding and workflow agents.