Xiaomi Drops MiMo Code V0.1.0: An MIT-Licensed CLI Coding Agent with Voice Input and Persistent Memory

What Is MiMo Code? / 什么是 MiMo Code？

On June 11, 2026, Xiaomi's MiMo team open-sourced MiMo Code V0.1.0, a terminal-native AI coding agent built as a fork of the OpenCode project. It ships under an MIT license — no strings attached. You start it by typing mimo in a terminal; the team recommends iTerm2 or VS Code's built-in terminal on macOS.

The release caps a wild week for Xiaomi's AI division. Three days earlier, on June 8, MiMo and TileRT demonstrated 1,000+ tokens per second decode on a 1-trillion-parameter MoE model running on a single 8-GPU commodity node — roughly 10x faster than the standard MiMo-V2.5-Pro. That same model family, from MIT-licensed MiMo-V2.5 and V2.5-Pro to Apache-2.0 MiMo-V2-Flash, now underpins a first-party coding agent CLI. Most Chinese AI labs ship models. Xiaomi just shipped the tooling too.

Why It Matters / 为什么重要

MiMo Code is Xiaomi's direct answer to Anthropic's Claude Code, and it enters a CLI coding agent market that barely existed 18 months ago. The differentiators are concrete:

• Voice input, for real. MiMo Code integrates MiMo-V2.5-ASR speech recognition. You can orally correct a typo or say "send" / "execute" to control the terminal. Claude Code has no equivalent.
• Memory architecture that doesn't rot. Instead of jamming everything into a single growing context window, MiMo Code runs a dedicated subagent that maintains three layers — project memory, session checkpoints, and task progress. When the context window fills, the subagent rebuilds a clean, compact summary. The philosophy, quoted from the IT之家 release: "主 agent 专心干活，记录完全外包" — the main agent focuses on work; recording is fully outsourced.
• The /dream command. An independent agent wakes every 7 days to merge, deduplicate, verify path validity, and compress scattered memories into a single compact current state. It's essentially automatic context hygiene.

Key Features / 核心功能

MiMo Harness + Compose Mode / MiMo Harness 与 Compose 模式

The MiMo Harness is a framework purpose-built for the MiMo model family. Hit Tab to enter Compose mode, which walks you through an end-to-end pipeline: design → planning → coding → testing → review. It's structured enough to prevent chaos, flexible enough to not feel like a wizard form.

Persistent Three-Layer Memory / 三层持久记忆

1. Project memory — what the agent knows about your codebase across sessions
2. Session checkpoints — snapshots within a single coding session
3. Task progress — where you are in the current task, so resumption is instant

The independent recording subagent means the main agent never context-switches to manage its own notes.

Voice Input & Control / 语音输入与控制

Powered by MiMo-V2.5-ASR. Correct typos orally. Say "execute" instead of hitting Enter. For developers who think out loud or prefer staying in flow, this is genuinely useful — not a gimmick.

Model Flexibility / 模型灵活性

MiMo Code ships with MiMo-V2.5 built in (native omnimodal: text, image, video, audio), but you're not locked in. Through the MiMo Token Plan, you can swap in DeepSeek, Kimi, GLM, and other third-party models. The Token Plan also works across Claude Code, OpenClaw, OpenCode, Kilo Code, Cline, Hermes Agent, and CodeBuddy Code — so credits aren't siloed to one tool.

The Broader MiMo Trifecta / MiMo 六月三连击

MiMo Code didn't land in a vacuum. Xiaomi's June 2026 can be read as a three-move sequence:

1. June 8 — UltraSpeed. MiMo-V2.5-Pro-UltraSpeed hits >1,000 tokens/s on a 1T-parameter MoE model using commodity GPUs. The secret sauce: FP4 quantization (MXFP4, MoE experts only) + DFlash speculative decoding (block-level masked parallel prediction, 6.30 acceptance length in coding tasks) + TileRT runtime. Three layers of codesign. 3x the price for ~10x the speed.

1. June 8–11 — Open-source model family. MiMo-V2.5 and V2.5-Pro (both MIT-licensed, 1M context window), plus MiMo-V2-Flash (Apache-2.0, on GitHub). Per InfoWorld, these are explicitly designed for "long-running AI agents" — the use case MiMo Code then embodies.

1. June 11 — MiMo Code V0.1.0. The CLI agent that ties the models to actual developer workflows.

How It Compares / 对比分析

MiMo Code's memory architecture is the most interesting design decision here. Offloading context management to a subagent means the main agent stays focused on code, and context compression happens proactively (every 7 days via /dream) rather than reactively when the window overflows.

Who Should Use It / 谁适合使用

• Developers already in the Xiaomi/MiMo ecosystem — native integration with MiMo-V2.5 models, Token Plan credits work across tools
• Mandarin-speaking developers — voice input trained on MiMo-V2.5-ASR, Chinese-language docs and community
• Anyone who wants an MIT-licensed Claude Code alternative — fork it, modify it, ship it
• Long-session coders — the persistent memory system is built for multi-hour coding sessions where context collapse is the usual failure mode

The Catch / 注意事项

V0.1.0 means exactly what it says. This is a first public release. Expect rough edges, missing integrations, and the kind of bugs that only surface when real developers hammer on a tool. The OpenCode foundation gives it a running start, but the gap between "promising CLI agent" and "daily driver" is measured in months of iteration. Whether Xiaomi sustains the open-source momentum — or treats this as a one-time PR move — will determine if MiMo Code matters a year from now.