MiMo Code: Xiaomi's open-source AI programming assistant challenges Claude Code.

The race among AI programming assistants is entering a new phase as Xiaomi officially introduces MiMo Code V0.1.0 – an AI programming tool that runs directly in the terminal and is released as open source. According to Xiaomi, this is not just a regular code writing assistant, but an agentic AI system designed to handle software development tasks that involve hundreds of consecutive steps, a weakness of many current tools.

Included with the MiMo Code is free, for a limited time, access to MiMo-V2.5 – Xiaomi's flagship multimodal AI model with a contextual window of up to one million tokens. Users can start using it without registering an account.

What is MiMo Code?

MiMo Code is an AI programming assistant that runs directly in the terminal environment, developed by the Xiaomi AI Team based on the open-source OpenCode project. Instead of building from scratch, Xiaomi focused on expanding this platform by adding long-term memory, specialized working modes, and state management mechanisms for complex software projects.

The project was announced on June 10, 2026, through MiMo's official account on the X social network. Xiaomi describes this product as "the smartest programming partner you've ever worked with," rather than simply an AI that assists in writing code.

Currently, MiMo Code is released under the MIT license and can be installed on macOS, Linux, or Windows via simple terminal commands.

The big problem for AI Coding Agents: Memory loss after long work sessions.

One of the most common problems with tools like Claude Code, Codex CLI, and many other coding agents is their ability to maintain context in long-running projects.

As the number of files, technical decisions, and development requirements increase, the model's context window gradually fills up. Critical information at the beginning of the project may be compressed or omitted, causing the AI ​​to forget the system's current conventions, architecture, and state. As a result, programmers have to constantly re-explain the project to the AI ​​every few hours or days.

According to Xiaomi, relying solely on contextual compression is not a long-term solution. Instead, a clear information storage and retrieval mechanism is needed to determine which data should be permanently stored and when it should be recalled.

MiMo Code's multi-layer memory system

The most notable feature of MiMo Code lies in its cross-session memory architecture.

Instead of relying solely on the context window of the AI ​​model, MiMo Code builds a storage system based on SQLite FTS5 with four different layers of information: project memory, session checkpoints, temporary notes, and progress logs for each task.

Specifically, Xiaomi deploys a secondary agent dedicated to recording project status. While the main agent focuses on programming, this agent continuously updates the system's memory with technical decisions, architectural changes, and work progress.

This model can be visualized as a large construction project. While the construction team is working on site, an architect continuously updates the drawings and design documents. When they need to return to an area that was completed a long time ago, the construction team doesn't have to recall every detail; they can simply refer to the previously stored documents.

Therefore, when the context starts to reach its limits, MiMo Code can recreate the working state from saved checkpoints instead of relying entirely on the model's short-term memory.

Two notable self-improvement mechanisms

In addition to long-term memory, MiMo Code is also equipped with two learning mechanisms based on usage.

The first feature is the /dream command , which runs periodically, about once a week. The system will review previous sessions, remove duplicate information, and consolidate it into a more organized long-term memory.

The second feature is distillation , which allows AI to analyze work history to identify frequently repeated processes. The system can then suggest or automate these processes in the future, similar to the approach OpenAI and Anthropic are taking with their latest agent models.

Performance on programming tests

According to data released by Xiaomi, MiMo Code, combined with the MiMo-V2.5-Pro ​​model, achieved higher results than Claude Code on several popular benchmarking tools for programming agents.

Interestingly, Xiaomi stated that the majority of the improvement doesn't come from the AI ​​model itself, but from the MiMo Code's agent architecture. When running the same MiMo-V2.5-Pro ​​model on two different frameworks, MiMo Code still achieved approximately 5 percentage points higher results on both SWE-bench Pro and Terminal Bench 2.

However, it's important to note that these figures are all self-published by Xiaomi and have not been independently verified. Furthermore, the company hasn't directly compared them to competitors like OpenAI's Codex CLI or Google's Gemini CLI.

In addition to traditional benchmarks, Xiaomi also conducted double-blind testing with 576 programmers across 474 private code repositories, generating over 1,200 direct comparisons with Claude Code.

The results showed that in short tasks under 200 execution steps, the two systems were almost equal. However, when the number of steps exceeded 200, MiMo Code's win rate increased to over 65%. This reinforces Xiaomi's hypothesis that long-term memory and state management are the platform's biggest advantages.

MiMo-V2.5: Xiaomi's strategic weapon.

MiMo Code is designed to integrate with a programmer's existing workflow rather than requiring them to change their habits.

The tool can read and write files, run terminal commands, manipulate Git, and supports direct import of MCP servers, customization, and API configuration from Claude Code. This makes the migration process relatively easy.

Another standout feature is Compose Mode . When this mode is activated, programmers only need to describe the goal at a high level. The AI ​​will automatically build the plan, design the solution, write the source code, test, and evaluate the results in a nearly self-contained process.

Additionally, MiMo Code supports voice control through MiMo-ASR voice recognition technology combined with the TenVAD voice activity detection system. Users can issue voice commands to create new requests or execute commands without using the keyboard.

One of the reasons MiMo Code is attracting attention is its accompanying AI model.

MiMo-V2.5 is a multimodal model with a Mixture-of-Experts architecture, possessing 310 billion parameters in total but only activating approximately 15 billion parameters in each inference. The more advanced version, MiMo-V2.5-Pro, has a scale of over one trillion parameters with 42 billion parameters active during the inference process.

According to Xiaomi, these models are trained to understand how modern agent frameworks work, including memory management and context in environments like OpenCode or Claude Code. Therefore, developing a separate framework optimized for MiMo is a logical step.

MiMo Code enters an increasingly crowded market already dominated by Claude Code, Codex CLI, Gemini CLI, OpenCode, and Aider. However, what's noteworthy isn't just the product itself, but also the entity behind it.

From a company renowned for smartphones and consumer electronics, Xiaomi is now investing heavily in AI with a series of models such as MiMo-7B, MiMo-VL, MiMo-V2-Pro, and MiMo-V2.5. This entire strategy is led by Luo Fuli – one of the key figures involved in the famous DeepSeek R1 project.

Xiaomi's approach is quite similar to DeepSeek or Qwen: releasing powerful models with open licenses and significantly lower prices than Western competitors to attract the developer community.

Should businesses care?

For engineering teams and businesses, MiMo Code is a worthwhile option to explore thanks to its MIT license, self-deployment capabilities, and support for using its own model instead of being required to connect to Xiaomi's infrastructure.

However, some points remain to consider. The system is currently at version 0.1.0, many performance metrics have not been independently verified, and the use of the default free service means data may be sent to Xiaomi's servers. This could be a barrier for organizations with stringent security or data storage requirements.

MiMo Code demonstrates that the AI ​​programming race is shifting from model competition to competition at the agent and framework levels. While many companies focus on scaling models, Xiaomi emphasizes memory management, maintaining long-term context, and optimizing the practical workflow for programmers.

While it's too early to say whether MiMo Code will truly surpass Claude Code or Codex CLI, the emergence of a new open-source competitor with long-term memory and free access to a powerful AI model will undoubtedly make the coding agent market more competitive in the near future.