GPN
GPU-NATIVE NETWORK
Download

Get GPN in one command

Install the node + miner/agent, generate a wallet, and start mining and serving AI jobs. Builds are published as rolling GitHub releases; the source repository is private today (access on request).

Private repository

The source repo and its release assets are private at this stage. The links below will return 404 until your GitHub account is granted access — ask the team to be added, then use them normally. Prebuilt artifacts do not require a Rust toolchain.

One-command installers

Two scripts cover the two install paths (a node, or an access point). Both run from a local clone of the repository (access required today) or from a hosted copy once the repo is public.

install.sh

Full node + miner/agent. Clones the source, builds gpn-node and gpn-agent, symlinks the binaries into ~/.local/bin, generates a wallet, installs a user-local Ollama + baseline models, and writes ~/.gpn/gpn.env.

./install.sh

Or pipe from a hosted copy (once the repo is public): curl -fsSL <REPO>/install.sh | sh. Override with GPN_PREFIX, GPN_BIN, GPN_NETWORK, GPN_REPO_URL.

install-accesspoint.sh

Run a public access point (gpn_scheduler) — the decentralized front door. A caller hits one URL and it routes each job to the best provider (least-ms idle node). Permissionless; no special trust, so running several is the point.

./install-accesspoint.sh

Add GPN_INSTALL_SERVICE=1 to install a systemd --user unit. Config lands in ~/.gpn/accesspoint.env — see the access-points page.

Release artifacts

Pre-built binaries are produced by CI and published as rolling releases — the tags always point at the newest green build, so download links stay stable. No build toolchain needed.

PlatformReleaseAssetContents
Windows windows-latest gpn-windows.zip gpn_node.exe, gpn_register.exe, provider binaries, Ollama installer scripts, run-windows-provider.ps1
Linux linux-latest gpn-linux.tar.gz gpn-node, gpn-agent binaries (provider, scheduler, GPU compute, …), example config, first-run notes
macOS macos-latest gpn-macos.tar.gz Same layout as Linux. Apple GPUs are not a mining target; serving is CPU/limited.

All three tags are private today. Windows: unpack into a bin\ folder next to run-windows-provider.ps1, then follow the Windows node path (it installs/starts Ollama for you). Linux/macOS: unpack and put the binaries on PATH, or just use install.sh.

Prerequisites

To run a node / provider

  • A discrete GPU — mining is GPU-only by construction.
  • OpenCL ICD for GPU mining — NVIDIA: nvidia-opencl-icd; AMD: rocm-opencl-runtime (then reboot).
  • Ollama for serving LLMs — required (the installer sets up a user-local copy). A node mines without it, but cannot serve AI jobs until it's installed and running.
  • Enough VRAM for your target models — an 8 GB card runs small quantized LLMs, SD-class image models, and LoRA fine-tunes; video/70B-class need a bigger card.

To build from source

  • Rust toolchain (rustup.rs) with the wasm32-unknown-unknown target.
  • Build deps: git build-essential clang llvm protobuf-compiler cmake pkg-config libssl-dev.
  • For a CUDA path, your vendor drivers/toolkit as needed.
  • Windows users serving jobs need no Rust/Python — the prebuilt provider drives Ollama.
No OpenCL detected?

Install the vendor OpenCL runtime above and reboot. The miner can fall back to CPU hashing, but at network difficulty a CPU is far below the GPU-only eligibility gate — a node needs a discrete GPU (or one meeting the measured-TFLOPS floor) to earn.

Ollama is required to serve LLMs

The provider talks to http://127.0.0.1:11434. Without ollama the node still mines, but every AI job fails (and gpn_node warns you at startup). Install + start it once:

./scripts/install-ollama.sh
# or:
curl -fsSL https://ollama.com/install.sh | sh
ollama serve &
ollama pull llama3.2:1b
powershell -ExecutionPolicy Bypass -File scripts\install-ollama.ps1
# or:
winget install Ollama.Ollama
ollama serve
ollama pull llama3.2:1b
AMD / Radeon (ROCm)

AMD is a first-class vendor: mining uses vendor-agnostic OpenCL (rocm-opencl-runtime, then reboot), and detection is automatic via rocm-smi/lspci on Linux or Win32_VideoController on Windows — so gpu_class/vram_mb/tflops are correct on AMD.

For serving LLMs, on Linux ollama runs on ROCm (RX 6000/7000 supported; if you hit "unsupported gfx" pin HSA_OVERRIDE_GFX_VERSION=11.0.0 for RX 7000 / 10.3.0 for RX 6000):

./scripts/install-ollama-amd.sh

On Windows, ollama's ROCm support is limited — an AMD box often falls back to the CPU (slow) or needs a DirectML build. Mining still works; just expect higher LLM latency. Prefer Linux for full AMD GPU serving.

Source

Clone it

git clone https://github.com/jeremyzguo/gpn_coin.git
cd gpn_coin
Open the repository Run-a-node guide → Next: deploy a node →