MiMo-Code — AI Coding Agent by Xiaomi Where Models and Agents Co-Evolve

Introduction

MiMo-Code is an open-source AI coding agent developed by Xiaomi's MiMo research team. It combines a purpose-trained code generation model with an agentic loop that can read files, run commands, and iteratively refine code changes in your repository.

What MiMo-Code Does

• Generates and edits code across multiple files in a repository
• Executes shell commands to test and validate changes in a feedback loop
• Understands project context by reading relevant source files on demand
• Supports multi-turn conversations for iterative refinement
• Co-evolves model and agent capabilities through reinforcement learning

Architecture Overview

MiMo-Code pairs a fine-tuned language model optimized for code understanding with a tool-calling agent runtime. The agent has access to file reading, writing, and command execution tools. A reinforcement learning pipeline continuously improves the model based on real coding task outcomes, creating a feedback loop between model quality and agent effectiveness.

Self-Hosting & Configuration

• Install the CLI via npm; the agent connects to local or remote model endpoints
• Configure model backend (local GPU, API endpoint) in the config file
• Set workspace root and file access permissions per project
• Customize tool availability (shell, file I/O, browser) through agent profiles
• Supports running behind corporate proxies with custom CA certificates

Key Features

• Co-evolution framework jointly improves the model and agent through RL
• Multi-file editing with automatic dependency tracking across changes
• Built-in test execution validates modifications before proposing them
• Conversation history persists across sessions for long-running tasks
• Extensible tool system allows adding custom capabilities

Comparison with Similar Tools

• OpenHands — broader AI dev agent with visual capabilities but no co-evolution training
• Aider — terminal-based pair programmer focused on git integration
• SWE-agent — research-focused agent for GitHub issue resolution
• Cline — VS Code extension with autonomous coding but IDE-dependent
• Continue — IDE code assistant focused on completion rather than agentic editing

FAQ

Q: Which programming languages does MiMo-Code support? A: The model is trained on Python, TypeScript, Java, Go, Rust, and C++, with best performance on Python and TypeScript.

Q: Can I use my own model backend? A: Yes. MiMo-Code supports OpenAI-compatible API endpoints, so you can point it at any compatible model server.

Q: How does the co-evolution training work? A: The model is periodically fine-tuned on successful agent trajectories, improving its ability to use tools and plan multi-step changes.

Q: Does it support private repositories? A: Yes. MiMo-Code runs locally and never sends source code to external servers when using a local model backend.

Sources

• https://github.com/XiaomiMiMo/MiMo-Code
• https://mimo.xiaomi.com/