Xiaomi has released MiMo Code, a new open-source coding agent that outperforms Anthropic's Claude Code on extended development tasks by using an innovative memory consolidation system. In 1,213 double-blind tests conducted by 576 real developers across 474 private repositories, MiMo Code won more than 65% of tasks once execution crossed 200 steps, according to independent benchmarking. The tool, released June 10 as an MIT-licensed fork of OpenCode, introduces a feature Xiaomi calls "Dream," which runs every seven days to compress and organize the agent's accumulated knowledge about a project.

What Makes MiMo Code's "Dream" Feature Different?

The core innovation behind MiMo Code's performance lies in its memory management system. Every seven days, while developers are not actively using the tool, an independent agent wakes up and performs a series of maintenance tasks on the coding agent's historical knowledge. This automated process reads through previous sessions, identifies and merges duplicate memories, removes references to files that no longer exist in the project, and compresses everything the agent knows about the codebase into a tighter, more efficient representation. This approach addresses a fundamental challenge in long-running AI coding sessions: as context accumulates, the agent's ability to reason about complex problems can degrade due to noise and redundancy in its memory.

The practical impact is significant for developers working on large, evolving codebases. Traditional coding assistants like Claude Code can struggle when tasks require more than 200 sequential steps because their working memory becomes cluttered with outdated or irrelevant information. By periodically consolidating this memory, MiMo Code maintains clarity and decision-making quality even in extended development sessions.

How Does MiMo Code Compare to Claude Code in Real-World Testing?

The benchmarking methodology behind MiMo Code's claims is noteworthy for its rigor. Developers inside Xiaomi ran 1,213 double-blind A/B tests, meaning neither the developers nor the evaluators knew which agent they were testing at any given moment. This approach eliminates bias and provides credible evidence of performance differences. The test cohort included 576 real developers working in 474 private repositories, ensuring the results reflect genuine development workflows rather than artificial benchmarks.

The results show a clear advantage for MiMo Code once tasks become sufficiently complex. While the source does not provide detailed performance metrics for shorter tasks, the 65% win rate on tasks exceeding 200 steps suggests that MiMo Code's memory consolidation strategy becomes increasingly valuable as development sessions grow longer and more intricate. One early user on Hacker News, where MiMo Code spent its release day on the front page, remarked that it was "hard to know if I was using Opus or MiMo blindfolded," comparing it favorably to Anthropic's Claude Opus model.

Steps to Deploy MiMo Code With Privacy Controls

• Download and Install: Clone the MiMo Code repository from its public source and follow the setup instructions in the README file to get the tool running on your local machine within approximately five minutes.
• Disable Default Telemetry: MiMo Code ships with telemetry enabled by default, with data sent to tracking.miui.com; review the configuration files and disable this feature if privacy is a concern for your development environment.
• Integrate With Your IDE: Configure MiMo Code to work with your preferred code editor or integrated development environment (IDE) by following the extension or plugin setup documentation provided in the repository.
• Test on a Sample Project: Run the agent on a smaller project first to understand its behavior and memory consolidation cycle before deploying it to critical production codebases.

What Does This Mean for the AI Coding Assistant Market?

MiMo Code's release reflects an emerging pattern in the terminal coding agent space: Anthropic introduces a new capability with Claude Code, and the open-source community responds with an alternative that addresses specific limitations. This competitive dynamic benefits developers by creating multiple options and pushing innovation forward. The fact that an open-source tool can match or exceed the performance of a proprietary model from a leading AI company suggests that the coding assistant market is maturing and that architectural innovations, not just raw model size, drive real-world performance gains.

The memory consolidation approach pioneered by MiMo Code could influence how other AI coding assistants handle long-running sessions. As development tasks become more complex and codebases grow larger, the ability to maintain clarity over extended interactions will become increasingly important. Developers evaluating coding assistants should now consider not just raw speed or initial accuracy, but how well a tool performs on multi-hour or multi-day development sessions that require sustained reasoning about a large codebase.

MiMo Code's open-source license and transparent engineering approach also lower barriers to adoption for teams concerned about vendor lock-in or data privacy. By running the tool locally and controlling telemetry settings, developers retain full visibility into how their code is processed, a significant advantage over cloud-based alternatives.