Xiaomi MiMo team officially open sourced on June 10MiMo Code V0.1.0——A terminal AI programming agent built based on OpenCode.Claude Code, and for the first time integrates persistent project memory, multi-agent collaboration and voice programming into an MIT open source license tool.
core competencies
The core architecture of MiMo Code is built around the three pillars of "computing, memory, and evolution", each of which directly addresses the pain points of current AI programming tools:
- Persistent memory across sessions: The bottom layer uses SQLite FTS5 full-text search engine to maintain a persistent
MEMORY.mdProject files. - Multi-agent collaborative architecture: Through the three major modes of analysis and planning (Max Mode, sampling 5 solutions in parallel and selecting the best), development execution, and task orchestration, the main and deputy agents can work in parallel.
- self-evolving system:
/dreamThe command automatically reviews historical sessions every week, removes duplicates and compresses them into long-term memory; the "distillation" function mines historical sessions for repetitive workflows that can be automated.ChatGPT Dreaming and AnthropicClaude Dreaming has the same purpose but different approaches. - Voice programming: Based on Xiaomi’s self-developed MiMo-ASR speech recognition technology, it supports voice command coding and architecture reconstruction, achieving true hands-free programming.
Benchmark performance
According to data released by Xiaomi’s technology blog, the MiMo Code paired with the MiMo-V2.5-Pro model surpassed all three standard evaluations.Claude Code (withClaude Sonnet 4.6):
- SWE-bench Verified:82% (vs. Claude Code 79%)
- SWE-bench Pro:62% (vs. 55%)
- Terminal Bench 2:73% (vs. 69%)
What is even more noteworthy is that an internal test covering 576 developers, 474 real private warehouses, and 1,213 sets of A/B comparisons showed:Within 200 steps, the two systems have a 50-50 win rate; after more than 200 steps, MiMo Code’s winning rate rises to over 65%——This is where the value of its persistent memory and state management architecture lies.
It should be pointed out that these are all self-reported data by manufacturers.
User experience/limitations
Advantages: One-click installation (one curl command for macOS/Linux, npm for Windows), zero-configuration out-of-the-box, automatic importClaude Code's MCP server and custom skills.DeepSeek,Kimi, GLM, etc.).
limitations: The V0.1.0 version number itself implies limited maturity; the "limited time" access to the free model means possible charges in the future, and the code context needs to go through Xiaomi servers; as a product of a Chinese technology enterprise, it may face the risk of compliance review by some organizations; the benchmark data has not been independently verified.
Overall Score
| 维度 | Score | evaluate |
|---|---|---|
| functional completeness | 8.0 / 10 | The three major innovations of persistent memory + multi-agent + voice are outstanding, but the V0.1 feature set is still being expanded. |
| 易用性 | 8.5 / 10 | One-click installation, zero configuration, automatic importClaude Code configuration; terminal native experience is smooth |
| Cost-effectiveness | 9.0 / 10 | MIT open source + free model access + extremely low API pricing, 96% cache hit rate significantly reduces costs |
| 中文支持 | 7.5 / 10 | Produced by the Xiaomi China team, Chinese can be understood naturally; documents and communities are mainly mixed in Chinese and English |
| 输出质量 | 8.0 / 10 | The stability of long-term tasks is outstanding, and SWE-bench performs well; however, the benchmark is self-reported and has not been independently verified. |
Overall rating: 8.2/10