Skip to main content

Native worker overview

Native workers run Qwen3.8 27B Uncensored on your machine via ollama — the network's model, registered as qwen3.8-27b-uncensored. They serve the 27B requests and earn 3-5x more than browser workers.

Why go native?​

  • Higher earnings — 27B jobs pay significantly more than browser jobs
  • Better model — 27B parameter model produces higher quality responses
  • Real GPU utilization — uses CUDA, Metal, or Vulkan for full hardware acceleration
  • Runs headless — no browser tab needed, runs as a background process or service

Requirements​

  • Node.js 18+ (22+ recommended)

  • ollama v0.32.15 or newer — the worker checks the version at startup and tells you to upgrade if it's older

  • One of:

    • NVIDIA GPU with 16GB+ VRAM (24GB recommended — RTX 3090, 4090, 5090)
    • AMD GPU with 24GB+ VRAM (RX 7900 XTX, via Vulkan)
    • Apple Silicon Mac with 32GB+ unified memory

    12GB cards and 16/24GB Macs are below the bar — they can't hold the model.

  • ~36GB free disk space — the downloaded weights are kept alongside ollama's copy (~20GB on macOS)

  • Stable internet connection

Quick start​

npx @compute-network/worker --token <your-token>

That's it. One command. The worker downloads the weights, builds the right model for your hardware, and detects your GPU automatically.

One model, no picker​

The network serves one public text model, so there is nothing to choose: every native worker runs Qwen3.8 27B Uncensored. The old --model flag is deprecated and ignored. Startup asks one question — Qwen worker (text) or Image worker — and remembers the answer; skip it with --mode max (the historical name for the text mode) or --mode image.

Which build you get​

The worker picks the build from the hardware it finds. There is nothing to configure:

HardwareBuildSpeculative decoding
NVIDIA, 24GB+GGUF Q4_K_MYes
NVIDIA, 16GBGGUF IQ4_XSYes
Several small NVIDIA cards, none big enough aloneone layer-split worker (noMTP build)No
AMD, 24GB+GGUF Q4_K_MNo
Apple Silicon, 32GB+GGUF noMTP Q4_K_M, on MetalNo (never on Metal)

Speculative decoding uses the model's own MTP head — it's lossless (same output, just fewer forward passes) and up to ~2.4x faster on code-heavy jobs.

GGUF weights are pulled from a pinned HuggingFace revision into ~/.config/compute-worker/models and kept there, so rebuilding the model doesn't re-download it. The context window is sized to your VRAM (8K-32K) and baked into the build.

Multi-GPU rigs​

More than one NVIDIA card? The worker detects them all and runs one worker per capable GPU automatically — no flags. Cards under 16GB are skipped, and if no single card can hold the model, the CLI runs one layer-split worker across the cards instead. Each card gets its own pinned ollama, logs are prefixed [gpu N], and --gpu 3 / --gpu 0,2,5 narrows it to the cards you choose. One prerequisite: stop any ollama already running on the box first, or GPU 0's worker adopts it and that daemon sees every card.

Full details, including the first-run download order and the 10-workers-per-IP cap: Linux setup → Multi-GPU rigs.

Get your token​

  1. Go to c0mpute.ai/earn
  2. Log in with your X (Twitter) account
  3. Scroll to Native Worker section
  4. Click Get Worker Token
  5. Copy and save the token — it's shown only once

See Worker tokens for more details.

Keeping the worker up to date​

The worker has no auto-update: it runs exactly the version you installed, and nothing self-upgrades at startup. Upgrading is an explicit step:

npm i -g @compute-network/worker@latest                       # global install
npx -y @compute-network/worker@latest --token <your-token>    # or pin @latest in your npx command

Every start prints its version (c0mpute worker v…), so you can always tell what a box is running.

Platform guides​

Image worker​

A native worker can run as an image worker instead of a text worker — it serves the image generation network by running ComfyUI + the Chroma1-HD model on your GPU, and earns per render.

npx @compute-network/worker --mode image --token <your-token>

On first run without --mode, the worker asks whether to run as a Qwen (text) worker or an Image worker and remembers your choice. Image mode downloads only the image model (~14GB), not the text model. A 24GB GPU (RTX 3090/4090) is recommended. Set COMFY_DIR if you want the worker to install/launch ComfyUI for you; otherwise point COMFY_URL at a ComfyUI you already run.