Xiaomi dropped MiMo Code on June 10, 2026, and within days it had 6,200+ stars on GitHub. That alone tells you something. But the star count is the least interesting part of this story.
MiMo Code is a terminal-native AI coding agent that scores 82% on SWE-bench Verified, beating Claude Code’s 79%. It ships with free access to MiMo V2.5 Pro, a 1-trillion parameter multimodal model with a 1-million token context window. It has a persistent memory system that no other coding agent offers. And it is completely open source under the MIT license.
This is the complete guide. We will cover what MiMo Code is, how it works, what makes it different, and whether it should replace your current setup.
What Is MiMo Code?
MiMo Code is a terminal-based AI coding agent built by Xiaomi. You run it in your terminal, point it at your codebase, and it can read, write, refactor, debug, and ship code autonomously.
If you have used Claude Code or OpenCode, the interaction model is familiar. You type natural language commands, and the agent executes multi-step coding tasks. The difference is in the details: the model powering it, the memory system, and the price (free).
MiMo Code is built on a fork of OpenCode, the open-source terminal coding tool. Xiaomi took that foundation and added their own model integration, a persistent memory layer, the /dream maintenance system, voice control, and workflow modes for different task types.
The project lives at github.com/XiaomiMiMo/MiMo-Code under the MIT license. You can fork it, modify it, self-host it, or contribute back.
Benchmark Performance
Here is where MiMo Code gets attention. The numbers are strong:
- SWE-bench Verified: 82% (Claude Code scores 79%)
- SWE-bench Pro: 62%
- Terminal-Bench 2: 73%
- Head-to-head: wins 65%+ of matchups past 200 execution steps
That last point matters most for real-world use. When tasks get complex and require many steps, MiMo Code pulls ahead. Short tasks are roughly equal across agents. But the longer and more involved a task becomes, the more MiMo Code’s architecture pays off.
The SWE-bench Verified score of 82% is currently the highest published score for any coding agent. Whether benchmarks translate directly to your daily workflow depends on what you build, but a 3-point lead over Claude Code on the most respected coding benchmark is not trivial.
For context on how these models compare more broadly, check our best AI coding tools for 2026 roundup.
The Model: MiMo V2.5 Pro
MiMo Code ships with free access to MiMo V2.5 Pro. This is Xiaomi’s flagship model:
- 1 trillion parameters
- 1 million token context window
- Multimodal (text, code, images)
- Hosted by Xiaomi, no API key purchase required
You can read our full breakdown of MiMo V2.5 Pro for model-specific details. The key point here is that you get a frontier-class model at zero cost. No subscription, no per-token billing, no usage caps announced as of launch.
If you want to use MiMo V2.5 Pro with other tools, we have a guide on setting it up with Claude Code as well.
Alternative Model Backends
MiMo Code is not locked to one model. You can swap in:
- DeepSeek (various models)
- Kimi (Moonshot AI)
- GLM (Zhipu AI’s models, including GLM-5.2)
This flexibility matters. If Xiaomi’s servers are slow in your region, switch to DeepSeek. If you want to test how GLM-5.2 handles your codebase, plug it in. The agent logic stays the same; only the model brain changes.
For a broader look at the Chinese AI model ecosystem, see our guide to the best Chinese AI models in 2026.
Persistent Memory: The Killer Feature
Most coding agents forget everything between sessions. You start fresh every time. MiMo Code does not.
It uses a SQLite database with FTS5 (full-text search) to store project context, decisions, file structures, and patterns across sessions. A background subagent continuously compresses your active context, extracting key information and storing it for future retrieval.
This means:
- Project conventions persist across sessions
- The agent remembers your architecture decisions
- File paths and relationships are retained
- Previous debugging context is available
When you return to a project after a week, MiMo Code already knows the codebase layout, your naming conventions, the tech stack, and what you were working on last. No re-explaining required.
We wrote a deep dive on MiMo Code’s persistent memory system if you want the technical details on how SQLite FTS5 works here and why it matters for long-running projects.
For comparison, Claude Code offers project-level memory through CLAUDE.md files, but that is static and manually maintained. MiMo Code’s memory is dynamic, automatic, and searchable. See our context window management guide for why this distinction matters.
The /dream Command
Every 7 days, MiMo Code can run its /dream command. This launches a maintenance agent that:
- Reviews all recent sessions
- Removes duplicate or outdated information
- Verifies that stored file paths still exist
- Compresses active memory into long-term storage
- Optimizes the memory database for faster retrieval
Think of it as garbage collection for your AI agent’s brain. Over time, without maintenance, memory fills up with stale references, deleted files, and redundant context. The /dream cycle keeps things clean.
This is genuinely novel. No other coding agent has anything like it. It solves a real problem: memory systems that grow without curation become noisy and slow. MiMo Code handles this automatically.
Installation
Getting MiMo Code running takes one command on Mac and Linux:
curl -fsSL https://mimo-code.xiaomi.com/install.sh | shOn Windows, use npm:
npm install -g @xiaomi/mimo-codeThat is it. No Docker, no complex dependency chains, no build steps. For the full walkthrough including model configuration and first-run tips, see our MiMo Code setup guide.
Voice Control
MiMo Code supports voice input. You can speak commands instead of typing them. This is useful for quick instructions while your hands are busy, or for developers who prefer verbal interaction.
The voice system works through your system microphone and processes locally before sending to the model. It is not a gimmick; it is a genuine accessibility and productivity feature.
Workflow Modes
MiMo Code offers different workflow modes optimized for specific task types:
- Code mode: Standard coding tasks, refactoring, feature implementation
- Debug mode: Focused on diagnosing and fixing issues
- Review mode: Code review and suggestions
- Plan mode: Architecture and planning discussions
Each mode adjusts the agent’s behavior, system prompts, and tool usage patterns for the task at hand. You can switch modes mid-session.
How It Compares
Here is the quick comparison with the major alternatives:
| Feature | MiMo Code | Claude Code | OpenCode |
|---|---|---|---|
| Price | Free | $20+/month | Free |
| SWE-bench Verified | 82% | 79% | Varies by model |
| Context window | 1M tokens | 200K tokens | Varies by model |
| Persistent memory | Yes (SQLite FTS5) | Static (CLAUDE.md) | No |
| Open source | Yes (MIT) | No | Yes (MIT) |
| Voice control | Yes | No | No |
For the detailed head-to-head, read MiMo Code vs Claude Code. For how it stacks up against the full field including ZCode, see MiMo Code vs ZCode vs Claude Code.
Who Should Use MiMo Code?
MiMo Code fits best if you:
- Want a powerful coding agent without a monthly subscription
- Work on long-running projects where memory persistence helps
- Prefer open-source tools you can inspect, modify, and self-host
- Are comfortable in the terminal
- Want flexibility to swap model backends
If you are already invested in the Anthropic ecosystem and value Claude’s writing quality for documentation, Claude Code remains strong. But purely on coding capability and cost, MiMo Code is the current leader.
Trust and Supply Chain Considerations
MiMo Code is open source and MIT licensed. You can audit every line. The model calls go to Xiaomi’s servers by default, which means your code is transmitted to their infrastructure for processing.
If you are working on proprietary code, consider whether sending it to Xiaomi’s servers aligns with your security policies. You can mitigate this by using alternative backends like DeepSeek or running local models, though that sacrifices the power of MiMo V2.5 Pro.
For a broader discussion of these concerns, read our AI model supply chain risks article.
What Comes Next
MiMo Code V0.1.0 shipped in June 2026. The roadmap includes:
- More model backend options
- Enhanced multi-file editing
- Team collaboration features
- Plugin system for custom tools
Given the pace of development and community interest (6,200+ stars in days), expect rapid iteration. The MIT license means the community can also drive features forward independently.
FAQ
Is MiMo Code really free?
Yes. As of launch, MiMo Code itself is free and open source, and access to MiMo V2.5 Pro (the default model) is free with no announced usage caps. This could change, but right now there is zero cost.
Can I use MiMo Code with models other than MiMo V2.5 Pro?
Yes. MiMo Code supports DeepSeek, Kimi, and GLM as alternative backends. You configure this in the settings. The agent works the same regardless of which model powers it.
How does MiMo Code compare to Aider?
Both are terminal-native and open source. MiMo Code has persistent memory and the /dream system, which Aider lacks. MiMo Code also ships with a free frontier model. Aider requires you to bring your own API key. See our Aider complete guide for more on that tool.
Is my code safe when using MiMo Code?
Your code is sent to whichever model backend you configure. With the default MiMo V2.5 Pro, code goes to Xiaomi’s servers. The agent itself is open source and auditable. Review your organization’s data policies before sending proprietary code to any external API.
Does MiMo Code work on Windows?
Yes. Install via npm (npm install -g @xiaomi/mimo-code). Mac and Linux use a curl one-liner instead. Full setup details in our installation guide.