Host a node

Linux, headless, one command.

The same node as the Windows app, without the window. It runs as a container on any Linux box with an NVIDIA or AMD GPU, keeps 80% of every job (85% on 16 GB+ cards), boosted up to 10x when the network is busy.

Docker, NVIDIA

Needs the NVIDIA driver and the NVIDIA Container Toolkit. 8 GB VRAM or more.

docker run -d --name bread-node --gpus all --restart unless-stopped \
  -v bread-node:/data \
  -e BREAD_WALLET=YourSolanaAddress \
  ghcr.io/breadai-io/node:latest

Docker, AMD (ROCm)

RX 6000 and 7000 series, and Instinct cards. The device flags replace --gpus.

docker run -d --name bread-node --device /dev/kfd --device /dev/dri --restart unless-stopped \
  -v bread-node:/data \
  -e BREAD_WALLET=YourSolanaAddress \
  ghcr.io/breadai-io/node:rocm

Then

The first start downloads the model (about 5 GB) into the volume. Restarts are instant.

docker logs -f bread-node     # model download, ONLINE, jobs and payouts
docker stop bread-node        # goes offline cleanly

Your node appears on the network page within a minute. Payouts land in your wallet on the hourly batch, each one with a Solscan link in the log.

Settings

BREAD_WALLET: required; the Solana address that receives payouts
BREAD_MODELS: bread-chat-8b (default), bread-vision-7b, or both, comma-separated
BREAD_NODE_ID: the name shown in Console; generated when empty
BREAD_REGION: US, CA, SA, EU, UK, AF, ME, IN, ASIA, AU, NZ, Other; from the time zone when empty
BREAD_MAX_JOBS: concurrent jobs, default 1; raise on 24 GB cards
BREAD_GPU: "NVIDIA RTX 3090,24" if detection fails

Without Docker

Already running Ollama? Download the tarball from the releases page, then:

tar -xzf bread-node-linux-x64.tar.gz && cd bread-node-linux
BREAD_WALLET=YourSolanaAddress node bread-node.mjs

Needs Node.js 22+. A systemd unit is included for running it as a service. Releases

What the container can see

The node makes one outbound connection to BREAD Console and talks to Ollama on the loopback interface. It receives prompts as text and returns text. Inside Docker it has access to the GPU device and its own volume, nothing else on the host. The image is built from the public Node-client repository, and you can build it yourself.

Windows app instead Host FAQ