Small enough to run on your Mac. Smart enough to ask.
小到能在你的 Mac 上跑，聪明到知道该问你。
Website · 官网 · Docs · 文档 · Download for Mac · Models · Discussions
DeskMind is open-source computer use for your Mac. It reads the screen, works out the next step and acts, with small models running on your Mac. When a task could mean two things, it asks instead of guessing.
- A small model, on your Mac. A 0.8B model decides each step and hands the unsure ones to a 4B, both locally with MLX.
- System One: choices, not guesses. Each step is a multiple-choice question with a probability for every option.
Unsure steps go to the 4B or to you. Any agent can use it through
/v1/systemone. - Open, from eyes to hands. Every part below is open source, along with the bench that grades it.
Start here: deskmind-ai/deskmind has the app, the demo and links to everything below. If DeskMind is useful or interesting to you, a star there helps other people find it.
| eyes | Sees the screen: finds the target when an app has no accessibility tree. |
| brain | Decides the next step: a probability for every option, a question when unsure. |
| hands | Acts on the desktop and checks the result before calling it done. |
| app | All of it in one Mac app (macOS 15+, Apple Silicon). Download: signed and notarized. |
| bench | Real-desktop tasks with graders that check the final state. |
| deskmind | Start here: the app, the demo, the roadmap and the launch post. |
Current release (October 2026): DeskMind 0.3.1 with Brain G18b. The models are on Hugging Face, mirrored on ModelScope for downloads from China; the app switches to the mirror by itself and checks every file's SHA-256. On the real macOS desktop (bench v25, 13 tasks × 3 runs, through the app on an M4 Pro), the 0.8B → 4B router at threshold 0.96 passed 39 of 39 runs and never said "done" early. Decision time: about 0.5 s when the 0.8B answers, about 3.6 s when the 4B answers (about 70% of steps). Small sample; what it still gets wrong is in brain results.
DeskMind 得心是在你 Mac 上运行的开源 Computer Use Agent：看屏幕、想下一步、动手操作，靠的是本机的小模型；任务有两种理解时， 它先问你，不瞎猜。
- 小模型，就在你的 Mac 上：0.8B 判断每一步，没把握的交给 4B 重新判断，都在本机用 MLX 运行。
- System One：选择，不是猜：每一步是一道选择题，每个选项都有概率；没把握就交给 4B 或先问你，任何 agent 都能通过
/v1/systemone接入。 - 从眼到手，全部开源：Eyes、Brain、Hands、Mac App 和给它们打分的 Bench。
从 deskmind-ai/deskmind 开始：App、演示和所有组件的入口都在那里。觉得有用或有意思，给它点个 star，能让更多人看到。
当前版本（2026 年 10 月）：DeskMind 0.3.1，内置 Brain G18b。下载 Mac 版 （已签名并经 Apple 公证）。模型在 Hugging Face 上，国内可从 ModelScope 镜像下载，应用会自动切换并逐个校验 SHA-256。在真实 macOS 桌面上（bench v25，13 个任务各跑 3 次，经 App 运行，M4 Pro）， 0.8B → 4B 路由（门槛 0.96）39 次全部通过，没有一次没做完就说完成。决策耗时：0.8B 自己回答约 0.5 秒，交给 4B 约 3.6 秒（约 70% 的步骤）。样本不大，还做不好的地方见 brain results。