Xiaomi has quietly dropped a new contender into the open-source AI arena. Their latest release, MiMo Code, is an agentic coding tool that claims to outperform Anthropic‘s Claude Code on complex, multi-step programming tasks. This Xiaomi open source coding agent is designed to handle long, intricate workflows where many other AI assistants stumble.
What sets MiMo Code apart is its focus on extended reasoning chains. Xiaomi says it excels on agentic coding benchmarks, particularly for tasks involving 200 or more steps. The tool is available under a permissive MIT license on GitHub, and to sweeten the deal, Xiaomi bundles limited-time free access to its flagship multimodal model, MiMo-V2.5. That model offers a million-token context window and requires no registration, making it practical for anyone wanting to test the open source coding agent without upfront costs.
How MiMo Code Beats Claude Code on Long-Horizon Tasks
That million-token context window gives MiMo-V2.5 impressive raw memory, but tackling complex, multi-step coding tasks requires more than just a large capacity. Agents often lose track of what they’ve done or need to revisit earlier decisions. That’s where Xiaomi open source coding agent MiMo Code pulls ahead of alternatives like Claude Code. Its architecture is built specifically for long-horizon coding tasks—work that spans many steps, sessions, or even days. Two key innovations make this possible: a layered memory system and a dedicated subagent that never interrupts your main workflow.
The Four-Layer Memory Architecture
MiMo Code’s cross-session memory uses SQLite FTS5 full-text search to keep its history instantly accessible. Instead of relying on a single conversation log, it organizes knowledge into four distinct layers:
- Project memory holds high-level goals and recurring patterns for each project.
- Session checkpoints capture snapshots of what was done during individual work sessions.
- Scratch notes let the agent jot down quick observations or ideas without interrupting the main task.
- Per-task progress logs track the status of each active step so nothing falls through the cracks.
This layered approach means the agent can pull relevant context from any prior session without sifting through an entire conversation. It also has a /dream command that reviews historical sessions roughly every seven days and compresses them into long-term memory, clearing out noise while preserving essential learnings.
The Role of the Checkpoint-Writer Subagent
Another reason MiMo Code excels at extended tasks is its design doesn’t force the primary agent to pause for note-taking. Instead, MiMo Code deploys an independent checkpoint-writer subagent that handles all the documentation and memory updates in the background. If you’re working on a debugging chain or refactoring several files at once, the main coding agent stays focused on writing and editing code. The checkpoint-writer silently logs progress, saves session checkpoints, and updates the progress logs. You get the benefit of full cross-session memory without the overhead of manual recalls or stalled workflows—exactly what you need for those projects that stretch across multiple work sessions.
Benchmark Performance: MiMo Code vs. Claude Code and OpenAI Codex CLI
That cross-session memory is a solid feature, but you’re probably wondering how MiMo Code stacks up against other coding agents in head-to-head benchmarks. Xiaomi has shared internal test results, and they show some interesting numbers. Keep in mind that these claims come directly from Xiaomi and haven’t been independently verified yet. With the Xiaomi open source coding initiative, MiMo Code enters a competitive field where raw performance on standard tests matters.
SWE-bench Verified Results
SWE-bench Verified is a key benchmark that tests how well coding agents handle real-world software engineering tasks, like fixing bugs and implementing features from actual GitHub issues. In Xiaomi’s internal tests, MiMo Code paired with MiMo-V2.5-Pro scored 82% on SWE-bench Verified. That beats Claude Code paired with Claude Sonnet 4.6, which scored 79%. It’s a narrow lead, but it places MiMo Code ahead of a major competitor on this metric.
Terminal Bench 2 Results
Terminal Bench 2 measures performance on command-line and terminal-based coding tasks, which is critical for developers who work extensively in the shell. When both agents used the same underlying model (MiMo-V2.5-Pro), MiMo Code scored 73% on Terminal Bench 2, while Claude Code scored 68%. That’s a five-point gap, suggesting MiMo Code’s agentic features—like its cross-session memory—add real value beyond the model itself.
Comparison with OpenAI Codex CLI
OpenAI’s Codex CLI is another major player, and the comparison here is more nuanced. On Terminal-Bench 2.0, Codex CLI running GPT-5.5 scores 82.2%, which is roughly nine points above MiMo Code’s self-reported 73%. So OpenAI leads on that terminal-specific test. But on SWE-Bench Pro, a harder variant of the benchmark, OpenAI reports GPT-5.5 at 58.6%, while MiMo Code plus MiMo-V2.5-Pro claims 62%. That gives Xiaomi an edge on more complex software engineering problems. The results paint a conflicting picture: no single agent dominates across all benchmarks, and your choice may depend on which tasks matter most in your workflow.
Open Source Roots and Customization Options
This is where the open-source nature of MiMo Code becomes particularly valuable. Since no single coding agent dominates every benchmark, your ability to tailor the tool to your specific tasks matters more than raw benchmark scores. MiMo Code is built on a foundation that you can inspect, modify, and extend — giving you control rather than a black box.
What is OpenCode?
MiMo Code is a fork of the open-source OpenCode agent. OpenCode is an agentic coding framework designed to be a flexible base for autonomous programming tasks. By starting from this existing framework, Xiaomi avoided reinventing the wheel and instead focused on adding its own improvements. This means you benefit from both the community-tested core of OpenCode and Xiaomi’s proprietary enhancements.
Xiaomi’s Extensions to the Base Agent
Xiaomi extended OpenCode with its own memory architecture, which helps the agent retain context across longer coding sessions. They also added specialized workflow modes that guide how the agent approaches different types of tasks, from bug fixing to feature implementation. A model harness ties everything together, allowing the agent to interact with underlying language models efficiently.
One of the most practical benefits of this open-source approach is customization. Because MiMo Code is an OpenCode fork, you are not forced to use Xiaomi’s model or configuration. You can potentially swap in other models that better suit your needs, whether that is a smaller model for faster iterations or a specialized model for a particular programming language. The agent’s architecture is designed to be modular, so you can adjust workflows or memory settings without starting from scratch.
Getting started is straightforward. On macOS and Linux, you can install MiMo Code with a single terminal command. Windows users can install it via npm. This simple installation process lowers the barrier to experimentation, letting you evaluate the agent quickly within your existing development environment.
For developers who value transparency and adaptability, the Xiaomi open source coding approach with MiMo Code offers a customizable alternative to proprietary agents. Whether you are a solo developer or part of a larger team, the ability to modify the agent’s behavior means you can align it with your specific workflow and coding standards.
Installation and System Requirements
Getting started with MiMo Code is refreshingly simple, especially given the complexity of what it does under the hood. The developers have kept the setup process lightweight, so you can move from download to your first coding task in just a few minutes. The Xiaomi open source coding agent runs on all major operating systems, and the installation steps are very straightforward.
Installation on macOS and Linux
If you are using macOS or Linux, you can install MiMo Code with a single terminal command. Open your terminal, paste the command provided in the official repository, and let the package manager handle the rest. The process is designed to be frictionless — no manual configuration, no dependency hunting. This makes it a practical choice for developers who want a macOS coding agent or a reliable Linux AI tool installation without the usual setup headaches. Once the command finishes, the agent is ready to run.
Installation on Windows via npm
Windows users can install MiMo Code through npm, the Node.js package manager. If you already have Node.js installed, the command is just as quick: open your command prompt or PowerShell, run the npm install command, and you are done. This approach keeps the setup consistent across platforms, so you do not need to learn a different process. After installation, you can start using the agent immediately.
One of the best parts? MiMo Code bundles limited-time free access to MiMo-V2.5, a multimodal flagship model with a million-token context window. That means you can work with very large codebases or long conversations without needing a separate registration or API key. The system requirements are minimal — a modern operating system and a stable internet connection are really all you need to get started. This low barrier to entry makes the Xiaomi open source coding tool accessible to a wide range of users, from hobbyists to professional teams.
Pricing, Limitations, and Future Access
Getting started with MiMo Code is remarkably straightforward, as you’ve seen. But what about the long-term picture? While the Xiaomi open source coding tool is released under the MIT license, meaning the code itself is freely available on GitHub, the free access to its flagship model is a temporary perk. This limited-time offer bundles MiMo-V2.5, a multimodal model with a million-token context window, requiring no registration. It’s a generous trial, but it raises practical questions about future costs and access.
What Happens After the Free Period?
As of now, Xiaomi has not announced any specific pricing for MiMo-V2.5 after the promotional period ends. The MIT license ensures that the base tool and code remain open-source and free, which is a strong foundation. However, the most advanced features—like the high-capacity context window and multimodal capabilities—may eventually come with a price tag. For now, you can enjoy this free AI coding tool without any financial commitment, but it’s wise to keep an eye on official announcements from the @XiaomiMiMo account for updates on MiMo Code pricing and future access. The release on June 10, 2026 via X marked the start of this promotion, so clarity on post-free access could arrive as the initial period winds down.
Developer Survey and Internal Testing
Xiaomi’s claim of beating Claude is based on its own internal beta release and a survey of 576 developers. While this suggests promising performance, it’s important to remember that these results come from a controlled setting. Real-world usage with diverse coding tasks and environments might yield different outcomes. The developer survey provides initial validation, but broader community testing will be key to understanding how MiMo Code truly stacks up against other tools. This data point is a starting point, not a definitive verdict.
Ultimately, the future of Xiaomi model access involves both an open-source core and potential premium tiers. For practical use, you can take advantage of the free period to evaluate MiMo-V2.5’s performance for your projects. If it meets your needs, you’ll have a clear incentive to follow Xiaomi’s pricing announcements. This combination of openness and temporary high-end access makes the Xiaomi open source coding tool a practical option to explore while it’s freely available, giving you a low-risk opportunity to test its capabilities firsthand.
Frequently Asked Questions
How does MiMo Code’s checkpoint-writer subagent improve coding efficiency?
The checkpoint-writer subagent saves incremental snapshots of your code at key intervals. This allows you to revert to a stable version quickly if a later edit causes errors. It reduces time spent debugging by providing clear rollback points during long coding sessions.
How does Xiaomi open source coding compare to OpenAI’s Codex CLI?
Xiaomi open source coding with MiMo Code focuses on handling extended tasks without losing context, while Codex CLI emphasizes rapid code generation. MiMo Code uses a specialized memory system to maintain coherence over long workflows. For sustained, multi-step coding projects, MiMo Code offers a practical advantage.
Is MiMo Code’s performance independently verified, or are the benchmarks only from Xiaomi?
Xiaomi has published internal benchmarks showing MiMo Code outperforming Claude Code on long coding tasks. Independent verification from third-party sources has not been widely released. You should test MiMo Code on your own projects to evaluate its real-world performance.
