Install on Linux

There are two ways to install: the prebuilt bundle (no compiler, no CUDA toolkit, only the NVIDIA driver) or building llama.cpp yourself. Both end with the same services.

Prebuilt bundle

You need an RTX 30, 40 or 50 series card with 24 GB, the NVIDIA driver 525.60.13 or newer (570.26 or newer for an RTX 50), an x86_64 Linux with systemd and glibc 2.35 or newer (Ubuntu 22.04, Debian 12 or newer), and about 20 GB of free disk space in /var/lib.

1. Download and check

base=https://github.com/aisathome/aisathome/releases/latest/download
curl -LO $base/aisathome-node-bundle-linux-amd64.tar.xz -LO $base/SHA256SUMS
sha256sum -c --ignore-missing SHA256SUMS
tar -xJf aisathome-node-bundle-linux-amd64.tar.xz && cd aisathome-node-bundle-linux-amd64

The bundle (about 0.5 GB) holds the installer, the node agent and llama-server with llama.cpp v0.5.0, built with CUDA 12.8 for the RTX 30, 40 and 50 series, plus the NVIDIA CUDA runtime and cuBLAS libraries it needs (see THIRD_PARTY_NOTICES.md). If the latest release has no bundle yet, use build it yourself.

2. Check, then install

./install.sh --dry-run                 # checks driver, GPU and disk; installs nothing
sudo ./install.sh

Paste the node token when asked (input is hidden). A token from your account page carries the node’s id; if we gave you a node id with your token, add --node-id <id>.

The installer then:

• checks the NVIDIA driver, the GPU and the free disk space, and stops with a clear message if one is not right;
• downloads the model (Qwen3.8 27B UD-Q4_K_M, 16.5 GB) from Hugging Face, pinned to one exact file, and checks its SHA-256. An interrupted download resumes when you run the installer again;
• writes node.json for the network’s model setting and the gateway wss://api.aisathome.org/v1/node/connect (--gateway-url for another);
• creates the aisathome user, the two services, a watchdog that restarts a hung llama-server, and a hook that stops the node around suspend, and waits until the model is loaded.

Add --enable-at-boot to start the node at boot. On a slow connection, download the model first without root: ./install.sh --download-only ~/models, then install with --model ~/models/Qwen3.8-27B-UD-Q4_K_M.gguf (the file is moved).

3. Check

journalctl -u aisathome-node -n 5

You should see connected to gateway. Then tell us, and we check that requests reach you.

Build it yourself

For other GPUs, other distributions, or if you prefer to compile. Tested on a fresh Ubuntu 22.04 machine with an RTX 3090 (driver 535) and on an RTX 4090.

1. Build tools and llama.cpp

llama-server must be built with CUDA. You need the CUDA toolkit that matches your driver (nvidia-smi shows the highest CUDA version it supports), cmake, ninja and a C++ compiler. On Ubuntu 22.04 with driver 535:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb && sudo apt-get update
sudo apt-get install -y cuda-toolkit-12-2 cmake ninja-build build-essential git
export PATH=/usr/local/cuda/bin:$PATH   # nvcc is not on PATH by default

git clone --depth 1 --branch v0.5.0 https://github.com/ggml-org/llama.cpp && cd llama.cpp
cmake -B build -G Ninja -DGGML_CUDA=ON "-DCMAKE_CUDA_ARCHITECTURES=86;89" -DCMAKE_BUILD_TYPE=Release
cmake --build build --target llama-server
cd ..

86;89 covers the RTX 30xx and 40xx series. A build for one of them only fails on the other with no kernel image is available for execution on the device.

2. The node agent and the installer

git clone https://github.com/aisathome/aisathome && cd aisathome

Until the first release is published, build the agent yourself. It needs Go 1.27 or newer from go.dev/dl (distribution packages are usually older):

go build -o aisathome-node ./node/cmd/aisathome-node

Once releases exist, --node-release latest downloads it instead and checks it against the release’s SHA256SUMS.

3. Install

sudo ./node/packaging/linux/install.sh --llama-build ../llama.cpp/build \
  --node-bin ./aisathome-node

As with the bundle, the installer downloads and checks the model, writes node.json and asks for the token. To use a model file you already have, add --model <file.gguf> (it is moved to /var/lib/aisathome-node/models/model.gguf). To write node.json yourself, pass --node-config node.json:

{
  "node_id": "your-node-name",
  "gateway_url": "wss://api.aisathome.org/v1/node/connect",
  "upstream": "http://127.0.0.1:8000",
  "models": [{ "name": "qwen3.8-27b", "roles": ["deep", "agent"], "slots": 2, "ctx": 32768 }]
}

Don’t put the token in the file: the installer asks for it. Add --dry-run to see the llama-server command without installing.