Xiaomi’s MiMo team has released and open-sourced MiMo Code, a terminal-based coding agent built for long-running programming tasks that can span dozens or hundreds of steps. The tool launched on June 11 as version 0.1.0 under the permissive MIT license, which lets developers use, modify and distribute it commercially. It is a fork of the open-source OpenCode project. Xiaomi says the agent’s main goal is keeping decision quality and state intact across very long sessions, where most coding assistants lose track of earlier work.
MiMo Code organizes its design around three ideas: computation, memory and evolution. Notable features include:
- A parallel reasoning mode called Max Mode that generates five candidate plans per turn and uses the model as a judge to pick the best, improving SWE-Bench Pro results by 10 to 20% at roughly four to five times the token cost.
- A completion verifier called Goal that checks a task against a user-defined stopping condition before the agent quits, cutting premature exits.
- A persistent, four-layer memory system that writes structured checkpoints to disk through a separate writer subagent, letting a session continue past the context window’s limit.
- Maintenance agents called Dream and Distill that run every 7 and 30 days to clean memory and turn repeated patterns into reusable skills.
- A Compose mode that takes an end goal and runs design, planning, coding, testing and review on its own.
The tool ships with one month of free access to Xiaomi’s MiMo-V2.5 model and a 1-million-token context, and it can also connect to third-party models such as DeepSeek, Kimi and GLM. Installation is a single shell command or an npm package. Its workflow engine is compatible with Anthropic’s Dynamic Workflow semantics and extends them.
Why Developers Care
- For long, automated jobs like migrating a whole codebase, the memory and orchestration features target the exact point where agents usually drift or stall.
- Open sourcing under MIT lets teams self-host and audit what the agent remembers, since memory is stored in editable files rather than an opaque database.
- The free model tier and multi-backend support lower the cost of trying it against existing tools.
Where It Fits
Xiaomi reports that MiMo Code paired with its MiMo-V2.5-Pro model beats Claude Code running Claude Sonnet 4.6 across three benchmarks, scoring about 62% on SWE-Bench Pro and 73% on Terminal Bench 2, a 5-point edge using the same base model. It also ran a double-blind A/B test across 576 developers and 474 private repositories, producing 1,213 scored comparisons. Win rates were near even on tasks under 200 steps but rose above 65% for MiMo Code once tasks passed 200 steps.
Xiaomi cautions that standard benchmarks still measure one-shot problem solving, so the multi-turn advantages mainly show up in real use. The release adds to a fast-moving open-source agent field and puts a major hardware maker directly against established tools from US AI labs.
