Run a node
A node is a computer with a strong NVIDIA GPU that serves the model to AIs@home users. It
runs two services: llama-server (the model, on 127.0.0.1 only) and aisathome-node, a
small agent that connects outward to the gateway and passes requests through.
Running a node is voluntary and unpaid. There is no token and no payment. Your node earns credit in BOINC’s unit: recognition, not money.
What you need
Section titled “What you need”- An NVIDIA GPU with 24 GB of memory: RTX 3090, 4090 or 5090 class.
- Linux with systemd and the NVIDIA driver (
nvidia-smiworks). - About 20 GB of disk for the model. With the prebuilt bundle, the install is mostly the 17 GB download; building llama.cpp yourself takes about an hour.
- A node token: sign in on your account page and choose Add a node (up to 3 per account). The network takes 10 nodes for now; when it is full, the page says so.
Windows is not supported yet.
Tested systems
Section titled “Tested systems”| GPU | System | Driver | Install | Result |
|---|---|---|---|---|
| RTX 4090 | Ubuntu 22.04.5 (desktop) | 530.30.02 | llama.cpp built by hand | The maintainer’s node, serving since 2026-10-09 |
| RTX 3090 | Ubuntu 22.04 (rented VM) | 535 | llama.cpp built by hand | Install and one chat through the gateway, 2026-10-10 (about 42 tokens/s) |
| RTX 3090 | Ubuntu 22.04 (rented VM) | 535.183.01 | prebuilt bundle | Install, download resume, connect, 1 × 128k in 18,302 MiB, about 42 tokens/s (2026-10-11). Needed a build fix: kernels must not be compressed for drivers before 570 |
| RTX 4090, RTX 5090 | rented | 595 | the bundle’s llama-server |
Runs; used for the model-check measurement (2026-10-11) |
Other 24 GB NVIDIA cards and distributions may work; the installer checks what it needs and says what is missing. Tell us what worked for you: contact@aisathome.org.
How the node runs
Section titled “How the node runs”- No open ports. The node only dials out; nothing listens on your network.
- Sandboxed. Both services run as the unprivileged user
aisathome: no access to/home, a read-only system, no new privileges. - Fewer traces on disk. No core dumps, no swap for the services, a bounded prompt cache in RAM, no prompt logging.
- Anonymous requests. Your node never learns who sent a request.
- Stop any time.
sudo systemctl stop aisathome-node aisathome-llamafrees the GPU.
How new nodes start
Section titled “How new nodes start”- A new node serves requests from day one. Requests with tools (agents such as opencode) reach it only from users who allow volunteer nodes for agents.
- Answers from volunteers’ nodes pass simple safety checks. If an answer looks like it injects a dangerous command, it is stopped and the node is held until we have looked at it; the account page shows that.
- No single volunteer gets more than about a third of the requests while others are free.
- Your node earns credit, except for requests you send yourself.
What you are responsible for
Section titled “What you are responsible for”- Your machine, its power bill and its other GPU work (BOINC, games). The node uses the GPU fully while it serves.
- Not looking at what you serve: don’t log, store, copy or inspect prompts or answers, and don’t tamper with them. This is a promise to users; see the terms.
- Keeping your node token secret, and telling us if it leaks.
Next: install on Linux.