MiMo-V2.5-Pro

Xiaomi has open-sourced MiMo-V2.5-Pro, a model with greatly improved agentic capabilities and long-horizon consistency. The model is built on a Mixture-of-Experts (MoE) architecture with a total of 1.02T parameters, with 42B active parameters, and is based on a hybrid attention architecture that supports a 1M token context window.

V2.5-Pro maintains high intelligence even in complex tasks requiring 1,000+ tool calls. In particular, it demonstrates strong consistency by adhering to fine-grained requirements even within long contexts.

Key achievements include:

• SysY compiler implementation: Built a complete compiler from scratch using Rust, covering everything from the Lexer to the RISC-V backend. Over 4.3 hours and 672 tool calls, it achieved a perfect score of 233/233 on the test suite.
• Full-featured video editor: Generated 8,192 lines of code from just a few prompts to complete a desktop app with multi-track timeline and audio mixing features. This task took 11.5 hours and 1,868 tool calls.
• Analog EDA design: Designed and optimized an FVF-LDO circuit based on the TSMC 180nm CMOS process. Using an ngspice simulation loop, it achieved all design targets in about 1 hour.

It is now immediately available through the API Platform and AI Studio, accessible by changing the model tag to mimo-v2.5-pro.