Xiaomi Open-Sources MiMo Code, an AI Agent That Beats Claude Code in Blind Test
Xiaomi has published a competitive open-source AI coding agent that, in its own testing, beats a well-known proprietary tool on both benchmarks and human preference.
Key Facts
- Xiaomi released MiMo Code on June 11, 2026, an open-source AI agent tool for the terminal.
- MiMo Code uses multiple agents to generate candidate execution plans and an evaluation agent to pick the best one.
- It converts natural-language skill files (SKILL.md) into deterministic JavaScript code to avoid inconsistent AI execution.
- In a blind test with 576 developers, MiMo-V2.5 achieved a 65% win rate over Claude Code on tasks exceeding 200 steps.
- The source code is published under the MIT License.
Reporting from 1 source: GIGAZINE.
Xiaomi released and open-sourced MiMo Code, an AI agent tool for the terminal that outperforms Claude Code in benchmarks and a blind test with 576 developers. It forks OpenCode and uses multiple agents to select the best execution path, plus deterministic JavaScript execution from natural language skill files.
Xiaomi released MiMo Code on June 11, 2026, an AI agent tool that runs on the terminal and is built from a fork of the open-source project OpenCode. Unlike single-agent systems, MiMo Code uses multiple agents to generate candidate execution plans and an evaluation agent to pick the best one. It also handles long-running tasks by extracting key points from sessions and starting new ones automatically.
MiMo Code introduces a method to convert natural-language skill files (SKILL.md) into deterministic JavaScript code, avoiding the inconsistent results that can occur when AI models execute instructions written in plain language. According to Xiaomi, MiMo Code recorded the highest scores in benchmarks when paired with the company's MiMo-V2.5-Pro model. In a blind test with 576 human developers, MiMo-V2.5 achieved a 65% win rate over Claude Code on complex tasks exceeding 200 steps. The source code is published under the MIT License.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.