About the binaries

Every release on GitHub is built by GitHub Actions from the tagged source, in public, by the workflow .github/workflows/release.yml. Nothing is built on anyone’s own computer.

What a release contains

File	What
aisathome-node-bundle-linux-amd64.tar.xz	Everything a Linux volunteer needs (about 0.5 GB): the installer, the node agent and llama-server with its libraries. See Install on Linux
aisathome-node-linux-amd64, -linux-arm64, -windows-amd64.exe	The node agent alone, for people who build llama.cpp themselves
aisathome-gateway-linux-amd64, -linux-arm64	The gateway, for people who run their own
SHA256SUMS	The SHA-256 of every file above

How the bundle is built

• The node agent and the gateway are Go programs built with CGO_ENABLED=0 (static binaries) by scripts/build-release.sh. aisathome-node -version prints the release.
• llama-server is llama.cpp v0.5.0, pinned to commit 7fe450e19305b828c199d602c23a8337aaa1f03b: the version our benchmarks measured. The build stops if the tag ever points elsewhere. scripts/build-llama.sh builds it inside NVIDIA’s own image nvidia/cuda:12.8.1-devel-ubuntu22.04, pinned by its digest, for the RTX 30 (sm_86), RTX 40 (sm_89) and RTX 50 (sm_120) series.
• No PTX is included, only code for those three GPU generations. That is what lets the CUDA 12.8 build run on older drivers (NVIDIA’s minor version compatibility); other GPUs need the build-it-yourself path.
• The bundle carries only the libraries llama-server actually loads (checked with ldd during the build, which fails on anything unexpected) and a BUILD-INFO file naming the llama.cpp tag and commit, the CUDA version and the GPU architectures.

Requirements and why

	Needed	Why
GPU	RTX 3090, 4090 or 5090 class, 24 GB	The model and its context need about 18 GB; the build targets these three generations
NVIDIA driver	525.60.13 or newer; 570.26 or newer for RTX 50	CUDA 12.x programs run on any 12.x driver from 525 on; the RTX 50 series needs CUDA 12.8’s driver
Linux	x86_64 with AVX2, systemd, glibc 2.35 or newer (Ubuntu 22.04, Debian 12 or newer)	The build image is Ubuntu 22.04
Disk	about 20 GB free in /var/lib	The model is 16.5 GB

The installer checks all of this before it changes anything (./install.sh --dry-run).

Checking a download

sha256sum -c --ignore-missing SHA256SUMS

This catches a corrupted or incomplete download. Be clear about its limit: SHA256SUMS comes from the same release page, so it does not protect against someone who takes over the release itself. Signed releases and build provenance are planned, not done yet. Until then, if you want certainty, build it yourself from the tagged source; the installer accepts your own llama.cpp build.

The model is downloaded by the installer from Hugging Face, pinned to one exact file and revision, and checked against its SHA-256 before use.

Licences

Part	Licence
AIs@home’s own code (installer, node agent, gateway)	Apache License 2.0
llama.cpp (llama-server and its libraries)	MIT, plus the licences in licenses/llama.cpp/
NVIDIA CUDA runtime and cuBLAS (libcudart, libcublas, libcublasLt)	NVIDIA’s CUDA Toolkit EULA, which allows shipping these files unmodified with an application. They are NVIDIA’s software, not covered by our licence
GNU OpenMP runtime (libgomp)	GPL-3.0 with the GCC Runtime Library Exception

Details are in THIRD_PARTY_NOTICES.md inside the bundle.

What has been tested

Releases are tested before we publish them, and we say plainly what was not. A test build of the first bundle (2026-10-11):

• RTX 3090, driver 535, Ubuntu 22.04: it failed at first: llama.cpp compresses its CUDA kernels by default, and drivers before 570 cannot load them. The build now leaves them uncompressed (a larger bundle). With that, the full path worked: install, an interrupted model download resumed, the node connected, 18.3 GB of GPU memory, about 42 tokens/s.
• RTX 4090 and RTX 5090, driver 595: the bundle’s llama-server ran without problems.
• Not tested yet: drivers between 525 and 535, and the release built by GitHub Actions (it is tested on a GPU before it is published).