Xiaomi MiMo Code: an Open-Source Coding Agent Gunning for Claude Code
Xiaomi just open-sourced MiMo Code, a terminal coding agent, and the target on its back could not be more explicit: 62% on SWE-Bench Pro, 73% on Terminal Bench 2, which Xiaomi says is about five points ahead of Claude Code. It hit the top of Hacker News today with 375 points. v0.1.0 shipped June 10 under MIT, built as a fork of OpenCode, written in TypeScript, 4.4k stars and climbing.
The genuinely novel part is how it handles memory. Instead of cramming everything into one context window, a background subagent continuously logs state, and when the window approaches its limit it saves a summary and hands the main agent a clean continuation. There is also a /dream function that runs weekly, reviewing past sessions, deduplicating and compressing what it learned. An agent that sleeps and consolidates memories β the biology metaphor is getting less metaphorical every month.
It ships with free access to MiMo-V2.5, Xiaomi's latest model, but you can swap in DeepSeek, Kimi or GLM as the backend. Voice input is built in, and a compose mode drives spec-to-tests workflows end to end. Install via curl or npm.
Take the benchmark claims with salt β the same day, Endor Labs published independent numbers showing Claude Fable 5 sitting mid-table on real vulnerability-fixing work. Vendor benchmarks are marketing until a third party reproduces them. But the strategy is unmistakable: like Kimi Work last week, Chinese labs are shipping aggressive free agent products as distribution for their models, and the open-source coding agent is now the beachhead everyone wants.
Repo: https://github.com/XiaomiMiMo/mimo-code
← Back to all articles
The genuinely novel part is how it handles memory. Instead of cramming everything into one context window, a background subagent continuously logs state, and when the window approaches its limit it saves a summary and hands the main agent a clean continuation. There is also a /dream function that runs weekly, reviewing past sessions, deduplicating and compressing what it learned. An agent that sleeps and consolidates memories β the biology metaphor is getting less metaphorical every month.
It ships with free access to MiMo-V2.5, Xiaomi's latest model, but you can swap in DeepSeek, Kimi or GLM as the backend. Voice input is built in, and a compose mode drives spec-to-tests workflows end to end. Install via curl or npm.
Take the benchmark claims with salt β the same day, Endor Labs published independent numbers showing Claude Fable 5 sitting mid-table on real vulnerability-fixing work. Vendor benchmarks are marketing until a third party reproduces them. But the strategy is unmistakable: like Kimi Work last week, Chinese labs are shipping aggressive free agent products as distribution for their models, and the open-source coding agent is now the beachhead everyone wants.
Repo: https://github.com/XiaomiMiMo/mimo-code
Comments