Xiaomi's MiMo Code Brings Persistent Memory to Terminal-Based AI Coding

Xiaomi has released an open-source terminal tool that remembers what you worked on yesterday. MiMo Code V0.1, built on the OpenCode project, adds long-term memory, self-improvement loops, and tight integration with the company's latest multimodal model. Developers can install it with a single command on Mac, Linux, or Windows, and it comes with free limited-time access to MiMo-V2.5 - a model boasting a one-million-token context window. That means entire codebases fit into a single prompt without losing critical details.

Most terminal assistants forget past sessions. MiMo Code does not. It automatically accumulates knowledge, using lossless compression to keep every line intact even as projects grow to a million lines. After each session, the system reviews what happened, distills lessons, and sharpens its recommendations. The more you use it, the smarter it gets.

In one test, a developer connected MiMo-V2.5 Pro to a compatible interface and let it run autonomously for 125 sessions. The model produced 301 git commits, generated over 60 pages of content - including landing pages, API docs, and a real-time pricing calculator - all for $70.12. A 96% cache hit rate kept repeated context costs near zero. Beyond the numbers, the model created its own product idea, built a backlog, ran quality audits, and fixed bugs it discovered without being asked.

Xiaomi built the agent framework around its own model, with a closed loop of testing, review, and validation handling complex jobs in one pass. Compose mode forces structure: specification to plan to build to report. Voice input, powered by MiMo-V2.5-ASR, supports hands-free prompts in multiple languages, including code-switching.

Compatibility is a strength. MiMo Code loads existing skills, MCP servers, and commands from Claude Code setups with zero migration effort. It also works with OpenAI, DeepSeek, Kimi, GLM, and other providers. The underlying MiMo-V2.5-Pro model uses a mixture-of-experts design with 1.02 trillion total parameters but activates only 42 billion. Its architecture includes hybrid attention, three-layer multi-token prediction, and a one-million-token context window. Xiaomi open-sourced the weights under MIT license on Hugging Face, allowing enterprises to self-host and researchers to fine-tune - though hardware demands are steep, with four A100 80GB cards as the minimum for local runs.

MiMo Code is part of a broader push. Earlier this year, Xiaomi released MiMo-V2-Pro in March, which quietly topped OpenRouter usage charts under the codename Hunter Alpha. The V2.5 series followed in April, adding native multimodal understanding of images, video, audio, and text. Daily token volume crossed one trillion in early April, and price cuts arrived in May.

Early reactions on X ranged from excitement to skepticism. Some called it "a fork of OpenCode with a different color theme," while others praised the free access to a strong model and persistent memory. Questions remain about long-term coherence and performance with third-party backends. Still, MiMo Code gives developers a coding partner that remembers yesterday's decisions and improves tomorrow's output - a compelling combination for teams chasing velocity on large codebases.

Source: Webpronews