Xiaomi's MiMo officially released and open-sourced a terminal AI coding assistant called MiMo Code V0.1.0 under the permissive MIT license. The tool is a spin on the open-source project OpenCode, comes with a limited-time free MiMo-V2.5 multimodal model built in, and supports APIs from DeepSeek, Kimi, and GLM, plus third-party token plans.
Most coding agents rely on the model taking notes, and they often forget key context if the model doesn't trigger an action. MiMo Code introduces a persistent memory system that outsources state recording to a separate subagent. When the conversation window gets near its limit, the subagent automatically saves the state and creates a clean summary for the main agent to pick up seamlessly. The built-in /dream command runs every 7 days: a separate agent reads through past chats and memory files, merges, deduplicates, and validates path integrity, compressing scattered info into a current state to update the global memory.
To improve the fit between the model and the agent framework, MiMo Code comes with a custom Harness system for the MiMo series of models. Users press Tab to switch to Compose mode, and after providing basic requirements, the system autonomously runs through a full development cycle: design, planning, coding, testing, and review. On the SWE-Bench Pro and Terminal Bench 2 benchmarks, using the same base model, MiMo Code scored 62% and 73% respectively, beating Claude Code by 5 percentage points.
MiMo Code also has built-in voice control, letting users modify input or send commands via voice. For deployment, macOS and Linux users can install with a single curl command, while Windows users can use npm. Once launched, the terminal shows a fully localized TUI interface with a status panel on the right to monitor progress.