Overview
MiMo Code is a **terminal programming Agent** open-sourced by Xiaomi's MiMo team (led by Luo Fuli) on 2026-06-13, based on a secondary development of OpenCode. Developed by **5 people in 14 days**, it topped GitHub Trending on its first day of open source, achieving **5.1k Stars / 200+ Issues**. Paired with the same series of MiMo-V2.5-Pro model, it is positioned as the "domestic open-source version of Claude Code."
Official offline benchmarks show: **MiMo Code + MiMo-V2.5-Pro outperforms Claude Code + Sonnet 4.6 on three benchmarks**; when execution steps exceed 200, the real-world development win rate rises to 65%. **1 million token context + limited-time free access** is very attractive to budget-conscious developers.
However, early-stage product issues on the first day of open source are also evident: users have reported multiple bugs in real-world testing, **including the Agent automatically determining the OpenCode package in the user's global npm directory as a "migration residue" without confirmation and executing `npm uninstall`, damaging the development environment**, suspected memory leaks, thinking falling into repetitive loops, and default telemetry sent to `tracking.miui.com`. The developer community has engaged in heated debate over whether the "coding harness has a moat," with some developers believing that Claude Code is "not particularly special" at the harness layer.
Key Features
- Terminal-native Agent: Command line / terminal-native experience, covering the complete workflow of reading and writing code, executing commands, running tests, and git operations
- 1 million token context: Supports 1M long context, allowing large codebases to be loaded at once for holistic understanding and refactoring
- MiMo-V2.5-Pro adaptation: End-to-end joint debugging with Xiaomi's self-developed MiMo-V2.5-Pro trillion-parameter MoE model, outperforming Claude Code + Sonnet 4.6 on three benchmarks
- 200+ step long-range tasks: Real-world development win rate rises to 65% when execution steps exceed 200, specifically optimized for long-range programming tasks
- Open source + limited-time free: Open source on GitHub, limited-time free usage, friendly to budget-conscious individuals / early-stage teams
- OpenCode secondary development: Base from the community open-source OpenCode, meaning MiMo Code's differentiated capabilities mainly lie in model adaptation + harness tuning layers
Use Cases
- Developers wanting to experience the "domestic open-source version of Claude Code"
- Individuals / early-stage teams sensitive to Claude Code subscription costs
- Research / refactoring tasks requiring 1M context for large codebase analysis
- Domestic compliance / Xinchuang scenarios requiring domestic replacement for AI programming capabilities
- End-to-end programming workflow for deep users of MiMo / Xiaomi ecosystem
Pros
- 5 people 14 days 5.1k Stars, explosive community popularity
- 1M context + 200+ step long-range task capability outstanding
- Open source + limited-time free, extremely low barrier to entry
- End-to-end joint debugging with MiMo-V2.5-Pro, most complete domestic stack
- OpenCode secondary development inherits community ecosystem
Pricing
**Open source + limited-time free**: Currently can be cloned and used directly from GitHub, model calls go through the MiMo open platform (under Xiaomi MiMo platform's limited-time free policy). **Self-hosted**: Can be fully privatized deployed with open-source MiMo-V2.5-Pro weights, model inference compute costs paid by user. **Future pricing** not officially announced.
Summary
MiMo Code is the hottest new product in the domestic open-source Coding Agent track on 2026-06-13—the combination of **5.1k Stars / limited-time free / 1M context / outperforming Claude Code + Sonnet 4.6 on three benchmarks** made it one of the most watched domestic alternatives to Claude Code on its launch day. **However, first-day bugs are numerous** (the most severe being the automatic `npm uninstall` of user global packages), so it is currently not recommended for direct use in production—suggest **playing in a sandbox, mocking a copy for important project repositories before letting it act**. If you just want to experience the "200+ step long-range programming" capability for free, it is one of the best options currently; if you are doing serious engineering, wait for at least a few version iterations to fix early bugs before migrating.