Xiaomi dropped an AI coding agent last week that hit 5,900+ GitHub stars in its first 48 hours. It beat Claude Code on SWE-Bench by 6 points, ships with a 1M token context window, and costs exactly $0.
No signup. No credit card. MIT license. Here's what's actually inside it and when it wins.
The Numbers That Matter
Instead of marketing fluff, let's look at what the benchmarks say:
| Metric | MiMo Code | Claude Code |
|---|---|---|
| SWE-Bench Verified | ~58% | ~52% |
| Context Window | 1M tokens | 200K tokens |
| Price | Free (limited time) | $20/mo + API |
| License | MIT (Open Source) | Proprietary |
| Real-world Win Rate (>200 steps) | 62% | 38% |
| Parallel Candidates | 5 (Max Mode) | 1 |
The stat that actually matters isn't SWE-Bench — it's the 62% win rate on tasks over 200 steps. When you throw a real, messy codebase at both tools, MiMo Code pulls ahead the longer the task runs. Claude Code wins on short, single-file fixes. MiMo Code wins on the kind of multi-file refactors and feature builds that eat your afternoon.
MiMo Code vs Claude Code: benchmark results and real-world task completion rates across different task lengths.
Why It Wins: Persistent Memory Isn't Just a Buzzword
Every AI coding tool claims "context awareness." Most just dump your last 100 messages into the prompt. MiMo Code does something different — it runs a SQLite FTS5-backed memory system that survives between sessions.
Here's what that means concretely:
Three Memory Layers That Actually Work
Project Memory (MEMORY.md) — The agent writes persistent facts about your codebase as it works. Architecture decisions, dependency quirks, unwritten conventions. When you open a new session tomorrow, it doesn't need to rediscover that your team uses ConvHandler instead of Handler for controller classes.
Checkpoint System (checkpoint.md) — A dedicated subagent (the "checkpoint-writer") watches the main agent's work and snapshots structured state at natural pause points. When context fills up — and with 1M tokens, that takes a while — it rebuilds from the last checkpoint instead of losing the thread.
Task Tree (tasks/<id>/progress.md) — Tasks get split into T1 → T1.1, T1.2 subtask trees. Progress is logged per leaf, so when a session resumes mid-task, the agent knows exactly what's done and what's left.
Here's the key difference: Claude Code has a 200K context ceiling. Once you hit it, you either /compact (losing detail) or start fresh (losing all context). MiMo Code's checkpoint system means the agent can keep working on the same task across multiple sessions without amnesia.
MiMo Code's three-layer persistent memory system: project memory, checkpoint snapshots, and task tree — all backed by SQLite FTS5 for fast context injection on session resume.
Getting Started in 30 Seconds
No configuration required if you use the free MiMo Auto tier:
# One-line install
curl -fsSL https://mimo.xiaomi.com/install | bash
# Or via npm
npm install -g @mimo-ai/cli
# Launch
mimo-code
First launch walks you through four options:
- MiMo Auto (free) — anonymous, zero config, uses MiMo-V2.5-Pro
- Xiaomi MiMo Platform — OAuth login if you want account features
- Import from Claude Code — migrates your Claude auth in one step
- Custom Provider — bring your own OpenAI-compatible API key
Pick option 1 and you're coding immediately.
The Secret Weapon: Max Mode
Add this to ~/.config/mimocode/mimocode.json or your project's .mimocode/mimocode.json:
{
"experimental": {
"maxMode": true
}
}
Max Mode runs 5 parallel candidate solutions at each reasoning step, picks the best via a judge model, and continues. It boosts SWE-Bench scores by 10-20% on complex tasks. The trade-off: it burns 5x the inference cost (but still $0 on the free tier while it lasts).
Three Agents, Three Jobs
MiMo Code splits work across three agent roles, switchable with Tab:
build — Full tool access. Reads files, writes code, runs shell commands, manages git. This is your default.
plan — Read-only. Explores your codebase, analyzes architecture, designs solutions. Use this when you want to think before executing.
compose — Orchestration mode. Built-in skills for specs-driven development: planning → TDD → implementation → code review → merge. Like having a senior engineer who follows a checklist.
Subagents spawn automatically as needed. If a task needs file reading while another subtask runs commands, both happen in parallel with lifecycle tracking.
The OpenCode Connection (And Why It Matters)
MiMo Code isn't built from scratch — it's a fork of OpenCode, the open-source coding agent from AnomalyCo. It keeps all of OpenCode's core capabilities:
- Multiple LLM provider support
- Terminal UI with Vim keybindings
- LSP integration for real-time diagnostics
- MCP server connections for tool extensibility
- Plugin system
What Xiaomi added on top:
- Persistent memory (the three-layer system above)
- Intelligent context management (checkpoints + budgeted injection)
- Subagent orchestration (parallel workers with lifecycle tracking)
-
Goal-driven loops (
/goalwith judge-model verification) -
Self-improvement (
/dreamextracts knowledge from sessions,/distillturns repeated workflows into reusable skills)
The fork relationship is a plus, not a minus. OpenCode's TUI and provider layer are battle-tested. Xiaomi focused their engineering on the parts that actually improve task completion rates: memory, planning, and autonomous execution loops.
When MiMo Code Wins (And When It Doesn't)
After testing on real projects, here's where each tool shines:
Use MiMo Code when:
- Building multi-file features across a large codebase
- Debugging bugs that span 3+ files and require understanding architecture
- Working on the same project day after day (memory compounds)
- You want an agent that doesn't stop halfway through a refactor
Use Claude Code when:
- You need a quick single-file fix or code review
- Working with Anthropic-specific APIs or MCP servers
- Tasks under 50 agent steps (roughly even with MiMo at this scale)
- You're already paying for Claude Pro and the $20/mo is sunk cost
The 200-step threshold is where MiMo Code's memory system creates separation. Below that, both tools are comparable. Above it, MiMo Code wins 62% of the time.
The Catch
There's always a catch. Here's what to watch for:
"Free for a limited time" — The MiMo Auto tier won't stay free forever. The MIT license means you can run the agent locally with your own API keys even if Xiaomi starts charging, but the free-inference gravy train has an expiry date.
It's 2 days old — GitHub creation date: June 10, 2026. 5,900+ stars in 48 hours is explosive, but production stability is unknown. Expect rough edges.
Chinese company, open-source code — Xiaomi is a Chinese hardware/software conglomerate. The code is MIT-licensed and auditable. The MiMo Auto service routes through Xiaomi servers for inference — use custom provider mode if you have API access concerns.
Windows support — Primary development targets macOS and Linux. WSL2 on Windows works; native Windows support is "coming soon."
What This Signals for AI Coding Tools
MiMo Code landing at #1 on Hacker News with 500+ points signals that developers are hungry for two things Claude Code isn't delivering:
Working memory across sessions — The single biggest pain point with current AI coders is re-teaching them your codebase every morning. MiMo Code's checkpoint system solves this.
Controlled autonomy — Claude Code stops and asks permission constantly. MiMo Code's
/goal+ judge-verification loop lets it work through multi-step tasks without babysitting while still having a safety check before declaring "done."
The open-source model also matters. Even if Xiaomi's free tier disappears, the code is MIT-licensed. You can plug in DeepSeek, GPT, Claude, or Ollama models and keep the memory/checkpoint/agent architecture — paying only for inference.
Bottom Line
MiMo Code isn't "Claude Code but free." It's Claude Code with a fundamentally different architecture — one that trades short-task parity for long-horizon superiority. If your AI coding workflow looks like "throw a 15-line fix at Claude and merge," stick with what you have. If you're trying to build features across sessions without spending 30% of your time re-prompting context, MiMo Code is worth the install.
The 1M token context window and five-way parallel Max Mode are impressive specs. The real story is the memory architecture underneath them.
Have you tried MiMo Code on a real project yet? What's your experience with persistent memory in AI coding tools — game changer or overhyped?
Top comments (0)