IA al Día
the efficient way to stay informed
Tools June 19, 2026 analysis 2 min read

Mimo Code: Xiaomi's Free Open-Source Terminal AI Agent Challenges Claude Code

MiMo Code is a completely free, MIT-licensed terminal AI coding agent from Xiaomi that beats Claude Code on long-horizon tasks and supports any LLM provider.

By IA al Día

If you caught our coding tools roundup, you already know MiMo Code made a splash — but this tool deserves a closer look. Released June 10 by Xiaomi’s MiMo team, MiMo Code is a terminal-native AI coding agent that’s wholly free (MIT license) and open source. No registration, no trial, no credit card. Just curl -fsSL https://mimo.xiaomi.com/install | bash and you’re off.

What makes it different

Built as a fork of OpenCode, MiMo Code adds capabilities absent from most terminal agents. Persistent project memory (backed by SQLite FTS5) remembers architecture decisions, rules, and progress across sessions — no more re-explaining context on resume. A subagent system spawns parallel workers for context reconstruction, checkpoint writing, and goal verification. Voice input is built in via TenVAD and MiMo ASR. The /dream and /distill commands let the agent self-improve: the former extracts knowledge from session traces into memory, the latter packages repeated workflows into reusable skills.

Benchmarks that turn heads

Xiaomi claims MiMo Code with MiMo-V2.5-Pro hits 62% on SWE-bench Pro and 73% on Terminal Bench 2 — edging out Claude Code with Sonnet 4.6 (57% and 68% respectively). The gap widens on 200+ step tasks, where Claude Code degrades. Roughly 5 points of the advantage come from the harness itself. For shorter tasks, the two trade blows evenly.

Any model, any provider

MiMo Code isn’t locked to Xiaomi’s models. It supports OpenRouter, any OpenAI-compatible API, and even one-click import of Claude Code authentication. The free “MiMo Auto” channel uses MiMo V2.5 (310B sparse MoE, 15B active, 1M token context). For heavier workloads, MiMo V2.5 Pro (1.02T total, 42B active) starts around $5/month on the MiMo platform.

Practical developer experience

Three agent modes — build (full permissions), plan (read-only analysis), and compose (spec-driven development) — are switchable via Tab. Task tracking uses a tree hierarchy (T1, T1.1, …) that persists with checkpoints. The /goal command sets a stopping condition with an independent judge model to prevent premature halts. Config resides in ~/.config/mimocode/mimocode.json or .mimocode/mimocode.json.

Caveats

It’s early days. The free tier has rate limits, macOS binaries can be flaky, some models are geo-restricted, and benchmarks are self-reported. Being a China-based product, data privacy deserves consideration. Still, given the zero-friction entry and MIT license, MiMo Code is worth a weekend spin alongside Codex CLI and Claude Code — especially if long-running autonomous tasks are your bottleneck.