Xiaomi open-sources MiMo Code terminal AI assistant with free multimodal model
Today, the Xiaomi technology team officially released and open-sourced an exploratory AI programming assistant that runs on the terminal — MiMo Code V0.1.0. This project is a derivative of the open-source project OpenCode and is released under the permissive MIT license, allowing individuals and companies to freely modify, distribute, and integrate it for commercial use.

As a self-evolving tool centered on "model-Agent collaborative optimization," MiMo Code aims to overcome the limitations of traditional AI coding tools. Its core highlights include:

Innovative persistent memory system that solves the "AI forgetfulness" problem: Traditional tools depend on the model's own recording capabilities, which often lose critical information in lengthy conversations. MiMo Code introduces three mechanisms—project memory, session checkpoints, and task progress—fully delegating the recording work to an independent Subagent. When the context window is full, the Subagent automatically rebuilds a clean summary, enabling the main Agent to continue seamlessly. Additionally, the built-in /dream command triggers automatically every 7 days to deduplicate, verify, and compress historical sessions, achieving memory consolidation and making the tool "understand you better over time."
Dedicated Harness system and Compose mode: To address compatibility issues between models and frameworks, Xiaomi has developed a dedicated Harness system for its built-in model. Users only need to press the Tab key to switch to Compose mode, enter a simple idea, and the AI automatically handles the entire workflow: "design, planning, coding, testing, and review." In real-world testing against Claude Code, MiMo Code adopts a strategy of "heavy upfront planning and stable later verification," which is slightly slower initially but produces more feature-rich, thoroughly tested industrial-grade products, resulting in less total time spent.
Excellent performance on authoritative test sets: Using the same MiMo model and relying solely on the Agent system's collaborative advantages, MiMo Code achieved a score of 62% on SWE-Bench Pro (beating Claude Code's 57%) and 73% on Terminal Bench2 (exceeding Claude Code's 68%).
Support for top-tier models and full Chinese localization: The tool includes a limited-time free multimodal model MiMo-V2.5 (with performance comparable to Claude Sonnet4.6), which is ready to use without registration. It also supports mainstream APIs and third-party token plans from providers like DeepSeek, Kimi, and GLM. A status dashboard is always displayed on the right side of the interface, and all settings have been fully localized in Chinese.
Voice interaction without manual operation: Leveraging MiMo-V2.5-ASR voice recognition technology, users can not only modify instructions verbally but also say commands like "send" and "execute" directly, achieving full hands-free control.
Blog: https://mimo.xiaomi.com/mimocode
Related article
Anthropic reveals rogue AI agents hate CAPTCHAs, just like you
Anthropic’s recent report on agentic misconduct highlights serious risks—its Mythos 5 model bypassed security controls to access the internet and upload malicious code to a public repository—but also reveals a humorous flaw: AI agents struggle with C
Tencent extends free access to Hunyuan Hy3, WorkBuddy and CodeBuddy users until August 5
Tencent has extended the free trial period for its Hy3 large language model. Following the conclusion of the initial two-week window, the company received overwhelming user feedback requesting continued access. In response, Tencent announced an exten
Shao Yin Debuts OpenFit 2 AI, Its First AI Open-Ear Headphones, With Deep Qianwen Integration for High-End Business Users
Having held the global market share lead in open-ear headphones for three straight years, Shan Yin has unveiled its inaugural AI hardware device: the OpenFit 2 AI. Leveraging Alibaba’s Qianwen large language model and tailored for high-stakes sectors
Related Special Topic Recommendations
Comments (0)
0/500
Today, the Xiaomi technology team officially released and open-sourced an exploratory AI programming assistant that runs on the terminal — MiMo Code V0.1.0. This project is a derivative of the open-source project OpenCode and is released under the permissive MIT license, allowing individuals and companies to freely modify, distribute, and integrate it for commercial use.

As a self-evolving tool centered on "model-Agent collaborative optimization," MiMo Code aims to overcome the limitations of traditional AI coding tools. Its core highlights include:

Innovative persistent memory system that solves the "AI forgetfulness" problem: Traditional tools depend on the model's own recording capabilities, which often lose critical information in lengthy conversations. MiMo Code introduces three mechanisms—project memory, session checkpoints, and task progress—fully delegating the recording work to an independent Subagent. When the context window is full, the Subagent automatically rebuilds a clean summary, enabling the main Agent to continue seamlessly. Additionally, the built-in /dream command triggers automatically every 7 days to deduplicate, verify, and compress historical sessions, achieving memory consolidation and making the tool "understand you better over time."
Dedicated Harness system and Compose mode: To address compatibility issues between models and frameworks, Xiaomi has developed a dedicated Harness system for its built-in model. Users only need to press the Tab key to switch to Compose mode, enter a simple idea, and the AI automatically handles the entire workflow: "design, planning, coding, testing, and review." In real-world testing against Claude Code, MiMo Code adopts a strategy of "heavy upfront planning and stable later verification," which is slightly slower initially but produces more feature-rich, thoroughly tested industrial-grade products, resulting in less total time spent.
Excellent performance on authoritative test sets: Using the same MiMo model and relying solely on the Agent system's collaborative advantages, MiMo Code achieved a score of 62% on SWE-Bench Pro (beating Claude Code's 57%) and 73% on Terminal Bench2 (exceeding Claude Code's 68%).
Support for top-tier models and full Chinese localization: The tool includes a limited-time free multimodal model MiMo-V2.5 (with performance comparable to Claude Sonnet4.6), which is ready to use without registration. It also supports mainstream APIs and third-party token plans from providers like DeepSeek, Kimi, and GLM. A status dashboard is always displayed on the right side of the interface, and all settings have been fully localized in Chinese.
Voice interaction without manual operation: Leveraging MiMo-V2.5-ASR voice recognition technology, users can not only modify instructions verbally but also say commands like "send" and "execute" directly, achieving full hands-free control.
Blog: https://mimo.xiaomi.com/mimocode
Anthropic reveals rogue AI agents hate CAPTCHAs, just like you
Anthropic’s recent report on agentic misconduct highlights serious risks—its Mythos 5 model bypassed security controls to access the internet and upload malicious code to a public repository—but also reveals a humorous flaw: AI agents struggle with C
Tencent extends free access to Hunyuan Hy3, WorkBuddy and CodeBuddy users until August 5
Tencent has extended the free trial period for its Hy3 large language model. Following the conclusion of the initial two-week window, the company received overwhelming user feedback requesting continued access. In response, Tencent announced an exten
Shao Yin Debuts OpenFit 2 AI, Its First AI Open-Ear Headphones, With Deep Qianwen Integration for High-End Business Users
Having held the global market share lead in open-ear headphones for three straight years, Shan Yin has unveiled its inaugural AI hardware device: the OpenFit 2 AI. Leveraging Alibaba’s Qianwen large language model and tailored for high-stakes sectors





Home






