Decision models (0.8B → 4B router) and the vision model behind the DeskMind Mac app 0.4.0. Same G18b release models as 0.3.
AI & ML interests
Computer-use agents, typed decisions, on-device (MLX)
Recent Activity
Organization Card
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 · Download for Mac · GitHub · Bench
DeskMind is open-source computer use for your Mac. It reads the screen, works out the next step and acts, with small models running locally with MLX. When a task could mean two things, it asks instead of guessing.
| Model | What it does |
|---|---|
| brain-0.8b | Fast tier: decides each step, with a probability for every option (MLX, 8-bit) |
| brain-4b | Strong tier: checks the steps the 0.8B is unsure of (MLX, 8-bit) |
| eyes-4b | Screen grounding: finds the target on a screenshot, for apps without an accessibility tree |
Current release: G18b (revision g18b-q8, also main). On the real macOS desktop (bench v25, 13 tasks × 3 runs
through the DeskMind app on an M4 Pro), the 0.8B → 4B router at threshold 0.96 passed 39 of 39 runs and never said
"done" early. A decision takes about 0.5 s when the 0.8B answers and about 3.6 s when the 4B checks. Small sample;
what it still gets wrong is in brain results. Any
agent can use the planner through /v1/systemone.
DeskMind 得心是在你 Mac 上运行的开源电脑操作 AI：看屏幕、想下一步、动手操作，靠的是本机的小模型（MLX）；任务有两种理解时， 它先问你，不瞎猜。
- brain-0.8b：判断每一步，每个选项都给出概率；brain-4b：复核 0.8B 没把握的步骤；eyes-4b：在截图上找到要点的位置， 用于没有辅助功能结构的应用。
- 当前版本 G18b：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。样本不大，还做不好的地方见 brain results。
- 国内下载：ModelScope 上有同样的文件（gxcsoccer/brain-0.8b、 gxcsoccer/brain-4b、 gxcsoccer/eyes-4b）。
datasets 0
None public yet