昨天刷到一条消息：小米开源了一个叫 MiMo Code 的终端编程助手。自报跑分超过 Claude Code，还限时免费。

跑分表到底能否信任？

小米给出的主要数字是：同一个 MiMo‑V2 模型，分别塞进 MiMo Code 和 Claude Code 两套壳子里跑。SWE‑Bench Pro 拿到 60% 对 58%，Terminal Bench 拿到 55% 对 53%。两边都领先大概 2 个百分点。

这类说法很聪明——它不是在宣称模型更强。而是在说“一样的模型，我的调度程序能多榨出一点点”。老实说，这样的对比是有意义的。

但存在一个大坑：这些分数全部在小米自己的机器上跑，MiMo Code 还没有出现在 SWE‑Bench Pro 和 Terminal Bench 的官方榜单上。自己关起门跑的数字，与别人按官方流程提交的数据之间会有配置差距。使用者往往会因为这样的不透明而产生不信任感。

装进终端先踩了个坑

安装非常快，只需一句：

npm install -g @mimo-ai/cli

十秒钟后即可在终端输入 mimo 启动。 它支持“一键导入”Claude Code 配置。但刚开始我直接用非交互模式执行 mimo run结果卡死整整几分钟，没有任何输出或报错。

痛点：免费通道需要先手动登录获取临时令牌，否则调用失败。

I had to run:

mimo providers login

并选择免费通道后再重试，才恢复正常。文档中没有明显提示这一步骤，新手很容易像我一样“撞个满怀”。

第一轮代码生成全翻车

Mimom Code 全程自行完成任务。大约五六分钟写完了 mdfm.js 和测试文件，并自行跑测试。结果这方面，

• Total cases: 10 PASSED: 6 BROKEN: 4

Pain point: The failure is subtle—tags are parsed incorrectly due to an overly greedy regex.

Poor error feedback makes debugging hard for beginners.

Mimom Code 自己挖坑又填好

The tool read failure report and immediately identified regex issue . It n fixed regex,re-ran tests,and all cases passed.

• No manual debugging required.
• A self‑repair loop: generate → test → diagnose → fix → retest.

This self‑repair loop is what many developers look for in an AI coding assistant—no more guessing where a bug lies.

Mimom Code 的设计灵感来自 Claude Code 与 Anthropic Dynamic Workflow

• A “Goal” mechanism ensures tasks only finish when all tests pass and code is committed—similar to Claude Code’s `/goal` command.
• A “Dynamic Workflow” lets main model generate deterministic scripts executed in a sandbox—explicitly compatible with Anthropic’s workflow semantics.
• The CLI itself is built on OpenCode,inheriting terminal UI,plugins。and protocol support.
• The real value lies in combining mature open‑source infrastructure with proven task design patterns from Claude Code,plus a free-to-use Chinese model and full localization.
• The touted “infinite context” is actually implemented by offloading state to files when window limits approach—a solid engineering trick but not true infinite memory.

User Pain Point: Expecting unlimited context can lead to disappointment when limits still exist under hood.

• If you want a free one‑million token context model or a fully Chinese terminal agent that can migrate from Claude Code smoothly—go ahead. The open source license allows experimentation without risk.

I’m not switching my primary tool yet. My publishing pipeline relies heavily on fine‑tuned Claude Code integrations . Switching bases would require substantial re‑engineering of dozens of tools—an unnecessary migration cost at this point.

The experience was different this time around compared with MiniMax M3: I saw it fail first hand n recover itself rar than just watching benchmark numbers on a chart.

Mimom Code shines on long tasks . For shorter tasks its win rate balances out 娱乐ween competitors;only when tasks grow longer does it outperform m significantly . My current test was too short to see this advantage fully.

Your main tool?其实，If you plan to evaluate a new programming agent on an actual project。what part do you fear will break?

Please share your thoughts in comments!