Xiaomi Launches Open-Source MiMo Code AI Assistant, Claims Superiority in Long Coding Tasks
Xiaomi's MiMo AI team has released MiMo Code V0.1.0, an open-source, terminal-native AI coding assistant. The company states it surpasses Anthropic's Claude Code on key agentic coding benchmarks, particularly for multi-step tasks exceeding 200 steps. MiMo Code introduces a cross-session memory system designed to prevent "AI coding agent amnesia" during extended development sessions. It is available on GitHub under an MIT license and comes with limited-time free access to Xiaomi's multimodal MiMo-V2.5 model, featuring a million-token context window.

Xiaomi's MiMo AI team has introduced MiMo Code V0.1.0, an open-source, terminal-native AI coding assistant. The Chinese electronics company asserts that MiMo Code outperforms Anthropic's Claude Code on significant agentic coding benchmarks, particularly for multi-step tasks requiring over 200 actions. This claim is based on internal beta testing and a survey involving 576 developers.
The new tool was announced on June 10, 2026, via the official @XiaomiMiMo account on X. MiMo Code is accessible on GitHub with an MIT license and supports installation across macOS, Linux, and Windows environments. It is based on the OpenCode agent, which Xiaomi has extended with its own memory architecture, workflow modes, and model harness.
A key feature of MiMo Code is its cross-session memory system, designed to prevent AI coding agents from losing context during long working sessions. This system employs four layers of memory—project memory, session checkpoints, scratch notes, and per-task progress logs—and utilizes an independent "checkpoint-writer" subagent to record progress. This allows the system to rebuild the environment from structured checkpoints with relevant context, maintaining operational momentum.
Xiaomi's internal benchmark figures indicate that MiMo Code, when paired with MiMo-V2.5-Pro, achieved higher scores than Claude Code with Claude Sonnet 4.6 across several evaluations. On SWE-bench Verified, MiMo Code scored 82% compared to Claude Code's 79%. For SWE-bench Pro, it scored 62% against 55%, and on Terminal Bench 2, it achieved 73% versus 69%. The company reported that the harness itself contributed approximately five percentage points to these gains.
An internal double-blind A/B evaluation with 576 developers found that while both systems performed similarly on tasks under 200 steps, MiMo Code's win rate exceeded 65% for tasks over 200 execution steps. This data supports Xiaomi's focus on its memory and state-management architecture for handling extended projects effectively.
MiMo Code integrates directly within the terminal, supporting file management, command execution, and Git operations without requiring extensive configuration. It provides limited-time free access to Xiaomi’s multimodal MiMo-V2.5 model, which features a million-token context window. The system also supports third-party backends and any OpenAI-compatible API, offering flexibility. Additional features include a "Compose mode" for specification-driven development and voice control for hands-free operation.
According to VentureBeat, Xiaomi's MiMo AI division, led by Fuli Luo, has been consistently expanding its AI offerings since April 2025, with MiMo Code being its latest significant release.
Advertisement
AdSense slot • inline
