opencompute worker

Native CLI worker for the opencompute contributor inference network. It connects to the orchestrator over Socket.io, serves Ollama-backed text jobs or ComfyUI-backed image jobs from your GPU, and accrues worker earnings for verified work.

Quick start

npx @opencompute/worker

The first run asks for mode/model, runs the local setup and benchmark, then opens a browser approval link and saves the worker session locally. Future starts can use the same bare command. For production, the CLI connects to https://api.openpute.com by default. Override it with OPENCOMPUTE_ORCHESTRATOR_URL or pass --url. For local development:

npx @opencompute/worker --url http://localhost:3001

On first run, the worker asks whether to run text or image mode. Text mode asks which Ollama model to run and remembers the choice. Interactive starts show live network counts when available and recommend the least-served text model. Headless starts reuse the saved mode/model, or fall back to text mode.

Models

The package currently supports these Ollama models:

• qwen3-8b - Qwen3 8B, fast general work, low VRAM (5GB).
• deepseek-r1-0528-qwen3-8b - DeepSeek R1 0528 Qwen3 8B, reasoning, low VRAM (5GB).
• qwen3-14b - Qwen3 14B, better quality, 16GB+ VRAM (9GB).
• dolphin-qwen-8b - Dolphin Qwen 8B, uncensored, low VRAM (5GB).
• dolphin-qwen-14b - Dolphin Qwen 14B, uncensored, 16GB+ VRAM (9GB).
• qwen3.6-35b-a3b - Qwen3.6 35B A3B MoE (3B active), 24GB+ VRAM (22GB).
• nemotron-3-nano-30b-a3b - Nemotron 3 Nano 30B A3B MoE (3B active), 24GB+ VRAM (20GB).

Skip the prompt with --model:

npx @opencompute/worker --model qwen3-14b

Runtime

Text mode checks for Ollama, installs it when possible, starts or restarts it with NVIDIA-friendly settings, pulls the selected base model, creates an opencompute model wrapper, runs a benchmark, then registers with the network.

Image mode checks for ComfyUI at COMFY_URL or manages a local install under ~/.opencompute. It pins ComfyUI to a tested commit, downloads the Chroma1-HD model files, runs a 512x512 render self-check, then registers as an opencompute-image worker.

Set OC_MANAGE_OLLAMA=0 if you want to manage Ollama yourself. The worker checks for package updates and prints the manual install command, but it does not install and re-exec updates unless OC_AUTO_UPDATE=1 is set. Set OC_NO_UPDATE=1 to skip the update check entirely.

Options

--url <url>      Orchestrator URL
--app-url <url>  opencompute web URL used for browser login
--mode <mode>    text or image
--model <model>  Model key to run
--benchmark      Run setup and benchmark, then exit
--version        Show version
--help           Show help

Earnings

Worker earnings accrue from completed verified jobs. The dashboard Rewards page shows pending, queued, and paid payout states, and the Wallet page controls the active payout wallet.

Image mode

npx @opencompute/worker --mode image

Useful environment variables:

COMFY_URL          Existing ComfyUI endpoint, default http://127.0.0.1:8188
COMFY_DIR          Existing ComfyUI folder, or unset for managed ~/.opencompute/ComfyUI
IMAGE_TIMEOUT_MS   Per-render timeout in milliseconds