Xiaomi has released MiMo Code V0.1.0, an open-source AI coding harness designed to work directly in the terminal. The company says the tool is built for agentic programming workflows and is especially aimed at long, multi-step development tasks that can stretch across multiple sessions.
The release was announced on June 10 by Xiaomi’s MiMo AI team and made available on GitHub under an MIT license. It can be installed on macOS and Linux with a terminal command, or on Windows through npm. Xiaomi also says the tool connects by default to a free, limited-time access channel for MiMo-V2.5, the company’s multimodal model with a million-token context window.
MiMo Code is based on the open-source OpenCode agent, but Xiaomi says it has added its own memory architecture, workflow modes and model harness. The company’s main pitch is that AI coding assistants struggle as sessions get longer, because older decisions and project details can fall out of context. Xiaomi says its system is designed to reduce that problem with structured storage and retrieval rather than relying only on context compression.
The company says MiMo Code uses a cross-session memory system built on SQLite FTS5 full-text search. That system spans four layers: a persistent project memory file, session checkpoints, scratch notes and task progress logs. Xiaomi says the goal is to preserve decisions and working state over time so the agent can recover its place after extended work.
A separate checkpoint-writer subagent handles note-taking so the main coding agent can keep working without interruption. Xiaomi describes this as a way to keep the primary agent focused on implementation while another process updates the project record in parallel. The system can then rebuild context from those structured checkpoints when needed.
The harness also includes two self-improvement features. A /dream command reviews past sessions on a roughly weekly basis, removes duplicate material and compresses useful information into long-term memory. A distill function looks for repeated workflows in prior sessions and turns them into reusable automation.
In its technical blog post, Xiaomi says MiMo Code paired with MiMo-V2.5-Pro outperformed Claude Code with Claude Sonnet 4.6 on three software engineering benchmarks: SWE-bench Verified, SWE-bench Pro and Terminal Bench 2. Xiaomi reported scores of 82% versus 79% on SWE-bench Verified, 62% versus 55% on SWE-bench Pro, and 73% versus 69% on Terminal Bench 2.
The company also said part of the improvement came from the harness itself. Using the same MiMo model in both systems, Xiaomi said MiMo Code still posted higher scores than Claude Code on SWE-bench Pro and Terminal Bench 2.
Beyond offline benchmarks, Xiaomi said it ran an internal double-blind A/B test with 576 developers working in 474 private repositories. The test produced 1,213 head-to-head comparisons against Claude Code. Xiaomi said the two tools performed about evenly on tasks under 200 execution steps, but MiMo Code’s win rate rose above 65% on work that went beyond 200 steps.
Those results have not been independently verified, and Xiaomi notes that standard benchmarks mainly measure one-shot problem solving rather than the kind of longer workflow management it is targeting.
Xiaomi says MiMo Code is intended to fit into existing developer routines. The tool runs in the terminal, can read and write files, execute commands and manage Git. It also offers a compose mode for more specification-driven workflows, and voice control for logged-in users.
The release arrives alongside Xiaomi’s broader push to make its MiMo models more accessible. The company is offering free, limited-time access to MiMo-V2.5 through MiMo Code, and it is positioning the software as a low-friction option for developers who want an agentic coding assistant with built-in memory management.
For now, Xiaomi’s message is clear: in its view, better agent performance may depend as much on the harness around the model as on the model itself.