How NanumOS works — your files, chats, calls and notifications
Where your files are
- Documents, Pictures, Movies and Music are inside NanumOS (the home folder) by default. In the Mac app you can link a folder of the Mac ("Connect a Mac folder"); it then shows as "Mac › name".
- Saving anything outside NanumOS on its own (a Mac folder, another device, a cloud) always asks first — an Allow card with the place and the size, every time.
- The home folder still holds up to 256 file and folder names (making it roomier is under way).
- Changed documents and settings save by themselves.
Where this chat is kept
- Your chats with Zeus are stored only in this device's home folder.
- A model on this device sends nothing out.
- Speaking and listening (dictation, the voice) run on each device.
Notifications
- The design for the Mac's Notification Centre is "without content" (who or what arrived, not the text).
Calls, messages and the exchange
- Installing NanumOS gives you a number; calls and messages go through your exchange. The exchange only connects outward (it opens no door in).
- Only people you know: you connect by invitation and approval on both sides. There is no public sign-up or number search — only people you invited (and approved) know your number.
- When the exchange runs on this computer and the computer is off, calls through that exchange stop too. How phones reach each other while it is off is not decided yet.
- The phone app has no VPN inside.
Without the internet
- Works: Documents, Sheet, Notepad, Calendar, Calculator, Photos, Retouch, the video editor, Music (songs on this device), the (English) Dictionary, Files, and Zeus when its model is on this device.
- Does not: listening to Radio, new News articles, Weather, the web, Zeus with an outside model, calls to others (unless on the same network).
The local model and the host's GPU
- The small model inside this device (the local model) normally answers on this computer's CPU. When NanumOS runs as a virtual computer on a Linux computer with a GPU (the host), it borrows the host's GPU and makes the same answer several times faster.
- Think of sending a heavy load on the neighbour's cart: you pack the load (the model) yourself, the cart (the GPU) is theirs. If the neighbour is using the cart (training), you do not borrow it and carry the load yourself (the CPU) — and it says so.
- The host gets only the model file's bytes and the questions. It cannot reach the host's files, folders or commands, and the host takes only a fixed few requests (hello, receive a model, make an answer, drop it, status).
- Try it now: ask Zeus to "check the system" — the "호스트 GPU" row shows when the GPU was last borrowed and how fast, or why it was not.
Technical details — the guest (NanumOS) gets one standard virtio-vsock device. llama --host-gpu sends the model with its sha256 to the host GPU service on vsock port 0x5a47 (a model the same guest sent before is not sent again) and shows the answer's pieces as they come. Zeus's local head (llama -serve) uses this path first when the channel device is there (not with LLAMA_HOST_GPU=0 in brain.env). The GPU driver stays on the host. The service refuses while a training program is on the GPU and takes an unused model off the GPU after 10 minutes. The last contact is kept in /tmp/hostgpu-<number> (never the question or the answer).
NanumOS 0.1.8349 · 2026-11-16