Xiaomi released MiMo Code V0.1 yesterday, an open-source terminal agent that remembers what it did last session. Built on OpenCode, it adds long-term memory, self-improvement loops, and tight coupling with Xiaomi’s MiMo-V2.5 multimodal model.
Installation is straightforward: one curl command on Mac/Linux, npm on Windows. It ships with limited free access to MiMo-V2.5, which offers a one-million-token context window. That means entire codebases fit in a single prompt without chunking or loss.
**Memory That Persists**
Most terminal assistants forget after you close the window. MiMo Code does not. It uses lossless compression to retain critical lines across sessions, even on million-line projects. After each session, it reviews what happened, distills lessons, and updates its internal best practices. The more you use it, the sharper it gets.
One tester connected MiMo-V2.5 Pro to a compatible interface and let it run autonomously for 125 sessions. It produced 301 git commits, generated over 60 pages of content (landing pages, API docs, a real-time pricing calculator across 33 models), and cost $70.12 for 387 million tokens. A 96% cache hit rate kept costs low—repeated context cost nearly nothing. The model even created its own product idea, backlog, and architecture, ran quality audits, and fixed bugs it discovered. Rate limits appeared during heavy runs but recovered quickly.
**Architecture and Workflow**
Xiaomi built the agent framework around its own model. A closed loop of testing, review, and validation handles complex tasks in one pass. Compose mode enforces structure: specification → plan → build → report. Developers think first, code second. Rework drops.
Voice input is supported via MiMo-V2.5-ASR, which handles multiple languages and code-switching. The tool loads existing skills, MCP servers, and commands from Claude Code setups—migration is seamless. It also works with OpenAI, DeepSeek, Kimi, GLM, and other providers.
**The Model Under the Hood**
MiMo-V2.5-Pro uses a mixture-of-experts design with 1.02 trillion total parameters, activating only 42 billion per forward pass. It features hybrid attention, three-layer multi-token prediction, and a one-million-token context window. Xiaomi released the weights under MIT license on Hugging Face. Enterprises can self-host, but hardware demands are steep: four A100 80GB cards for comfortable local runs. Most users will stick with the API.
**Context and Trade-offs**
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 (images, video, audio, text). Daily token volume crossed one trillion in early April. Price cuts arrived in May. Older V2-Pro and Omni endpoints will be deprecated at the end of this month.
MiMo Code is part of this broader push. The GitHub repository (XiaomiMiMo/MiMo-Code) contains full source under MIT. Early reactions on X were mixed—some called it “a fork of OpenCode with a different color theme,” while others praised the free model access and persistent memory.
**Open Questions**
Can the memory system stay coherent over months-long projects? How does performance hold when switching from MiMo-V2.5 to third-party backends? Edge cases in complex architecture decisions need broader testing against frontier models.
Still, the move fits Xiaomi’s pattern: open weights, agentic workloads, long context, low cost. MiMo Code brings those strengths into the daily terminal workflow. For teams chasing velocity on large codebases, a coding partner that remembers yesterday’s decisions and improves tomorrow’s output carries immediate appeal.
Installation is simple. The MIT license invites forks and commercial integration. As competitors tighten rate limits and raise prices, an open tool backed by cheap, high-context inference stands out. Whether MiMo Code becomes the default terminal agent or accelerates the next wave of forks, it has already shifted the conversation about what long-horizon coding agents should look like.
Source: Webpronews