Xiaomi open-sourced a coding agent that beats Claude Code — by forking the tool everyone else uses

A coding agent, not a model

The submitted link was in Chinese, so the thread's most-upvoted comment did the translating. A regular who pointed everyone to the English repo summarized it: "MiMoCode is a terminal-native AI coding assistant. It can read and write code, run commands, manage Git, and use a persistent memory system… while continuously improving itself." The code lives on GitHub under an MIT license.

The benchmarks are real and good: 62% on SWE-Bench Pro and 73% on Terminal Bench 2, roughly five points ahead of Claude Code using the same underlying model, with Xiaomi's post-training explicitly teaching the model "harness awareness" — how to manage its own memory and context inside an agent scaffold. The pitch is long-horizon work: it's reported to beat Claude Code on ultra-long, 200-plus-step tasks.

The README line that set the tone

What the announcement led with, the thread led with too, except as a critique. An ex-Meta engineer who'd worked on HHVM quoted the giveaway straight from GitHub:

"MiMoCode is built as a fork of OpenCode. It keeps all core OpenCode capabilities… and adds persistent memory, intelligent context management, subagent orchestration, goal-driven autonomous loops… and self-improvement via dream/distill."

That framing produced the obvious objection. A commenter asked: "Why not just contribute to OpenCode instead of creating a clone?" — and another just squinted: "Looks an awful lot like OpenCode."

"Open source" — but which kind?

The word did a lot of lifting, and someone checked the load. A commenter asked the question that decides whether this is a tool or a funnel:

"Is that Open-Source like, run it locally, no phone home included, or open source like the thin front-end layer is all that is actually open-source but it's an empty shell without the remote API it relies on?"

The practical answer leaned positive, with caveats. A developer who'd switched to it that day, routing it through a third-party plan, reported it "very pleasant so far, nice UI and seems to respond faster than Claude Code," while noting the snappy MiMo-2.5-pro model "does cost more than DeepSeek, though, so I'm not sure whether it's worth it yet." Another would-be tester got geofenced out of the fast model entirely.

The real story is the harness, not the hype

Strip the marketing and MiMo Code is a credible bet that the next gains come from the scaffold around the model — memory, context discipline, self-distillation — not the weights alone. That it's an OpenCode fork is less an indictment than the point: the agent layer is now competitive enough to fork and tune. Xiaomi's contribution is the training that makes a model fluent in its own harness. Whether "open source" survives contact with a geofenced default model is the question the thread left open.