Xiaomi releases open source MiMo Code AI assistant, outperforms Claude on long coding tasks

Xiaomi MiMo Code Challenges Claude Code with Superior Long-Horizon Performance

Xiaomi's MiMo AI team announced the release of Xiaomi MiMo Code V0.1.0 on June 10, 2026, positioning this open source terminal-native AI coding assistant as a formidable Claude Code competitor. According to internal benchmarks and a survey involving 576 developers, the agentic AI coding harness outperforms Anthropic's Claude Code specifically on long-horizon, multi-step coding tasks exceeding 200 steps 1

1

. Available now on GitHub under an MIT license, the tool installs with a single terminal command on macOS and Linux, or via npm on Windows 1

1

. The release bundles limited-time free access to MiMo-V2.5, Xiaomi's multimodal flagship model featuring a one-million-token context window that requires no registration to begin using 1

1

.

Cross-Session Memory Tackles AI Coding Agents' Amnesia Problem

The distinguishing feature of this AI coding assistant lies in its persistent project memory architecture, which addresses a critical weakness plaguing existing coding agents. As context windows fill during extended sessions, earlier decisions and conventions typically get compressed or lost entirely, forcing developers to repeatedly re-explain their projects. Xiaomi's approach deploys a cross-session memory system powered by SQLite FTS5 full-text search spanning four layers: project memory stored in a persistent file, session checkpoints, scratch notes, and per-task progress logs 1

1

. Rather than pausing the primary coding agent to take notes, the system employs an independent checkpoint-writer subagent that updates structured state in real time 1

1

. When context limits approach, the system rebuilds the environment from these checkpoints with relevant context, maintaining operational momentum without information loss 2

2

.

Self-Improvement Mechanisms and Multi-Agent Architecture

Two self-improvement mechanisms enhance the system's capabilities over time. A "dream" command periodically reviews historical sessions roughly every seven days, deduplicating and compressing them into long-term memory 1

1

. The "distill" function mines past sessions for repeated workflows that can be automated, following approaches recently adopted by OpenAI and Anthropic 1

1

. The multi-agent system supports build, plan, and compose modes, with the compose mode following a structured workflow from specifications through planning, execution, debugging, testing, verification, and merging 2

2

. Subagents can be created for parallel tasks with full lifecycle management and tree-structured task tracking 2

2

.

Benchmark Results Show Gains on SWE-Bench and Terminal Bench 2

According to figures published by Xiaomi, MiMo Code paired with MiMo-V2.5-Pro achieved 82% on SWE-bench Verified versus Claude Code with Claude Sonnet 4.6 at 79%, 62% versus 55% on SWE-bench Pro, and 73% versus 69% on Terminal Bench 2 1

1

. The harness itself contributes measurably to performance: running the same MiMo-V2.5-Pro model in both harnesses, MiMo Code scored 62% on SWE-bench Pro versus 57% for Claude Code, attributing roughly five points purely to the agent system 1

1

. Xiaomi conducted a human double-blind A/B evaluation during internal beta with 576 developers working in 474 real private repositories, producing 1,213 judged head-to-head pairs against Claude Code 1

1

. However, comparisons against OpenAI's Codex or Google's Gemini CLI were notably absent, with the official Terminal-Bench 2.0 leaderboard showing OpenAI's Codex CLI running GPT-5.5 at 82.2%, roughly nine points above MiMo Code's self-reported 73% 1

1

.

Broad Compatibility and Installation Options

Built as a fork of the open-source OpenCode agent, MiMo Code extends it with memory architecture, workflow modes, and model harness capabilities 1

1

. The system supports multiple model providers including Anthropic, OpenAI, DeepSeek, Kimi, and GLM, with Claude Code compatibility that automatically loads skills, MCP servers, and commands 2

2

. Additional features include voice input powered by MiMo-V2.5-ASR for speech-based prompts and an experimental Max Mode for parallel best-of-N reasoning with judge selection 2

2

. Developers can configure onboarding through MiMo Auto with zero configuration and anonymous access, Xiaomi MiMo Platform via OAuth login, one-step migration from Claude Code, or custom OpenAI-compatible API providers 2

2

.