Overview
MiMo-V2.5-Pro is the flagship reasoning model of Xiaomi's MiMo series, publicly tested on April 23, 2026, and officially open-sourced on April 28, 2026 (MIT license). It is the first model from a phone manufacturer to top the global open-source model rankings. With a total of 1T parameters and 42B activated MoE architecture, along with a 1 million token ultra-long context, it focuses on complex logical reasoning, multi-step task planning, autonomous programming, and other core Agent scenarios.
Key Features
- 1T MoE Architecture: Total parameters of 1 trillion, 42B activated, with mixed attention improving long-text efficiency
- 1 Million Token Context: Supports input at the level of entire books or entire code repositories
- Claw-Eval Score 75.7: Top three globally, second in China, benchmark score for logical reasoning
- AAII Score 49: Artificial Analysis Intelligence Index: second in China, eighth globally
- Programming Capability Comparable to Opus 4.6: Cost only 20%, outstanding cost-effectiveness
- MIT Open-Source License: Weights already on Hugging Face, zero licensing barriers for commercial use
Use Cases
- Teams needing Claude Opus-level programming capability but with budget constraints
- Main control brain for Agent workflows (multi-step planning + tool calling)
- Ultra-long document analysis, whole-repository code understanding
- Flagship model replacement under domestic compliance requirements
Pros
- Cost-effectiveness is top 1 among current domestic flagship open-source models
- MIT license is more permissive than Apache 2.0, zero barriers for redistribution
- 1 million token context is first-tier in the open-source camp
- Already supported by mainstream Agent frameworks such as Claude Code, OpenClaw, and Cline
Pricing
API pricing: ≤256K context input $1.00/M, output $3.00/M; ≤1M context input $2.00/M, output $6.00/M. The open-source version can be self-deployed, with weights available on Hugging Face. New users registering at platform.xiaomimimo.com receive ¥10 trial credit.
Summary
MiMo-V2.5-Pro is one of the most significant releases in the domestic open-source model landscape in 2026. The combination of 'programming comparable to Opus + price only 20% + MIT open-source' provides Agent developers with a cost-effective alternative to Claude Opus. It is recommended for teams working on Agent or programming workflows to evaluate it first.