Run AI inference on the Dria network. Earn rewards by serving models from your machine.

Choose one installation method:

Homebrew (macOS / Linux):

brew install firstbatchxyz/dkn/dria-node
dria-node --version

Homebrew will add the tap automatically.

Shell script (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/firstbatchxyz/dkn-compute-node/master/install.sh | sh
dria-node --version

AMD ROCm (Linux x86_64):

curl -fsSL https://raw.githubusercontent.com/firstbatchxyz/dkn-compute-node/master/install-rocm.sh | bash
dria-node --version

Requires ROCm 6.x to already be installed on your machine.

PowerShell (Windows):

irm https://raw.githubusercontent.com/firstbatchxyz/dkn-compute-node/master/install.ps1 | iex
dria-node --version

From GitHub Releases:

Download the latest file for your platform from Releases, then run dria-node --version to verify it.

Run the interactive setup:

dria-node setup

This will:

1. Detect your system RAM and list models that fit
2. Let you pick a model from the available options
3. Download the GGUF model file from HuggingFace
4. Run a test inference to verify everything works
5. Print a benchmark (tokens per second)

Use --gpu-layers -1 to offload all layers to GPU (Metal on macOS, CUDA on NVIDIA builds, ROCm on AMD Linux builds):

dria-node setup --gpu-layers -1

Once setup is complete, start the node:

dria-node start --wallet <YOUR_SECRET_KEY> --model <MODEL_NAME>

The node will connect to the Dria network, register your models, and start serving inference requests. You can increase throughput with --max-concurrent:

dria-node start --wallet <KEY> --model lfm2.5:1.2b --max-concurrent 4

Serve multiple models by comma-separating them: --model "qwen3.5:9b,lfm2.5:1.2b"

Override quantization with --quant Q8_0 (applies to all models).

dria-node <COMMAND>

Commands:
  setup    Interactive setup: pick a model, download it, and run a test
  start    Start the compute node

setup options:
  --data-dir <PATH>        Data directory [env: DRIA_DATA_DIR]
  --gpu-layers <N>         GPU layers to offload (0 = CPU only) [default: 0]

start options:
  --wallet <KEY>           Wallet secret key, hex-encoded [env: DRIA_WALLET]
  --model <MODELS>         Model(s) to serve, comma-separated [env: DRIA_MODELS]
  --router-url <URL>       Router URL [default: quic.dria.co:4001] [env: DRIA_ROUTER_URL]
  --gpu-layers <N>         GPU layers to offload (-1 = all, 0 = CPU) [default: 0]
  --max-concurrent <N>     Max concurrent inference requests [default: 1]
  --data-dir <PATH>        Data directory [env: DRIA_DATA_DIR]
  --quant <QUANT>          Override GGUF quantization [env: DRIA_QUANT]
  --insecure               Skip TLS verification [env: DRIA_INSECURE]

All flags can also be set via environment variables.

git clone https://github.com/firstbatchxyz/dkn-compute-node.git
cd dkn-compute-node
cargo build --release

Feature flags:

• --features metal — Apple Metal GPU acceleration (macOS)
• --features cuda — NVIDIA CUDA GPU acceleration
• --features rocm — AMD ROCm GPU acceleration (Linux x86_64)

cargo test

cargo clippy
cargo fmt --check

This project is licensed under the Apache License 2.0.