Xiaomi just dropped a new open-source tool that turns the terminal into a persistent coding partner. MiMo Code V0.1 arrived yesterday. It builds directly on the OpenCode project yet adds layers of long-term memory, self-improvement, and tight integration with the company’s latest multimodal model.
Developers can install it with one command. On Mac and Linux, a simple curl script gets it running. Windows users grab it through npm. The package comes paired with free limited-time access to MiMo-V2.5. That model packs a one-million-token context window. Suddenly entire codebases fit inside a single prompt without losing details.
Memory That Lasts Across Sessions
Most terminal coding assistants forget what they did yesterday. MiMo Code does not. It accumulates knowledge automatically. Lossless compression keeps every critical line intact even when projects swell to a million lines. The system reviews each session afterward. It distills lessons and best practices. The more a developer uses it, the sharper the agent becomes.
This self-evolving loop marks a clear break from one-shot tools. One tester connected MiMo-V2.5 Pro to a compatible interface and let it run autonomously. Over 125 sessions the model produced 301 git commits. It generated more than 60 pages of content including landing pages, API docs, and a real-time pricing calculator across 33 models. Total cost came to $70.12 for 387 million tokens. A 96% cache hit rate made that number possible. Repeated context cost almost nothing.
But the numbers only tell part of the story. The model created its own product idea, backlog, and architecture. It ran quality audits without prompting and fixed bugs it discovered. For structured tasks like static sites or serverless functions the performance felt reliable. Rate limits appeared during heavy runs yet recovered after short waits.
Xiaomi built the agent framework around its own model. A closed loop of testing, review, and validation handles complex jobs in one pass. Compose mode forces structure. The workflow moves from specification to plan to build to report. Developers think first and code second. Rework drops.
Voice input adds another dimension. MiMo-V2.5-ASR powers hands-free prompts. Speak the request. The system turns speech into action. Support for multiple languages and dialects including code-switching makes the feature practical for global teams.
Compatibility matters too. MiMo Code loads existing skills, MCP servers, and commands from Claude Code setups. Migration takes zero effort. The tool also works with OpenAI, DeepSeek, Kimi, GLM and other providers. Flexibility stays high.
The underlying model deserves attention. MiMo-V2.5-Pro counts 1.02 trillion total parameters yet activates only 42 billion in a mixture-of-experts design. Hybrid attention, three-layer multi-token prediction, and one-million-token context define its architecture. Xiaomi open-sourced the weights under MIT license on Hugging Face. Enterprises can self-host. Researchers can fine-tune. Hardware demands run steep. Four A100 80GB cards mark the minimum for comfortable local runs. Most users will stick with the API.
Earlier this year Xiaomi released MiMo-V2-Pro in March. It quietly topped usage charts on OpenRouter under the codename Hunter Alpha before the official launch. The company followed with V2.5 series in April. Native multimodal understanding of images, video, audio and text pushed agent capabilities higher. Daily token volume crossed one trillion in early April. Price cuts arrived in May. Deprecation of older V2-Pro and Omni endpoints takes effect at the end of this month.
MiMo Code sits inside this bigger push. The official page frames it as an exploratory terminal assistant. A companion blog post titled “MiMo Code: Scaling Coding Agents to Long-Horizon Tasks” explains the design choices (Xiaomi MiMo Blog, June 10, 2026). The GitHub repository XiaomiMiMo/MiMo-Code contains the full source under MIT terms.
News outlets moved fast. YugaTech reported the release this morning and noted claims that MiMo Code outperforms Claude Code on certain benchmarks. AASTOCKS highlighted the on-device aspects and open-source nature for Xiaomi’s Hong Kong-listed shares. The-Decoder covered the broader V2.5-Pro release in May, positioning it against Claude Opus on autonomous coding tasks (The Decoder).
Developers on X reacted with a mix of excitement and skepticism. One post called it “a fork of OpenCode with a different color theme.” Others praised the free access to a strong model and the persistent memory system. A detailed thread outlined the architecture: infinite context through compression, agent-model synergy, self-evolving reviews, voice input, and broad provider support. The official XiaomiMiMo account announced the launch late on June 10 and saw thousands of engagements within hours.
Questions remain about real-world scale. Can the memory system stay coherent over months-long projects? How does performance hold when the model switches from MiMo-V2.5 to third-party backends? Edge cases in complex architecture decisions still need broader testing against the latest frontier models.
Still, the move fits Xiaomi’s pattern. The company has released multiple models in quick succession. Many carry open weights. Focus stays on agentic workloads, long context, and low cost. MiMo Code brings those strengths into the daily workflow of terminal users. It gives developers a coding partner that remembers yesterday’s decisions and improves tomorrow’s output. For teams chasing velocity on large codebases, that combination carries immediate appeal.
Installation remains straightforward. The project page offers exact commands and troubleshooting steps. Early adopters already share tips on running it inside iTerm or VS Code terminals for best results. The MIT license invites forks, commercial integration, and further experimentation.
Xiaomi’s timing feels deliberate. 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 simply accelerates the next wave of forks, it has already shifted the conversation about what long-horizon coding agents should look like.
WebProNews is an iEntry Publication