Skip to main content

Windows setup

Two options: WSL (recommended) or native Windows. Either way you need an NVIDIA GPU with 16GB+ VRAM (24GB recommended), ~36GB free disk, and ollama v0.32.15 or newer — the worker checks the version at startup and stops if it's older. Grab the current build from ollama.com/download (or, inside WSL, curl -fsSL https://ollama.com/install.sh | sh).

WSL gives you a Linux environment inside Windows with better CUDA support and fewer compatibility issues.

Install WSL2​

wsl --install

Restart your computer after installation.

Install Node.js in WSL​

Open your WSL terminal (Ubuntu) and run:

curl -fsSL https://deb.nodesource.com/setup_22.x | sudo -E bash -
sudo apt install -y nodejs

Install CUDA in WSL​

NVIDIA GPUs are automatically passed through to WSL2. You need the CUDA toolkit inside WSL (not the Windows version):

  1. Make sure you have the latest NVIDIA Game Ready or Studio driver on Windows
  2. In WSL, follow the NVIDIA CUDA on WSL guide
  3. Verify:
nvidia-smi    # Should show your GPU
nvcc --version # Should show CUDA version

Run the worker​

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

There is nothing to choose: every native worker serves the same model, qwen3.8-27b-uncensored. The old --model flag is deprecated and ignored. Startup asks whether to run a Qwen worker (text) or an Image worker — skip it with --mode max (the historical name for the text mode):

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

On first run the weights download from a pinned HuggingFace revision and the worker builds the model for your card (Q4_K_M with speculative decoding on 24GB+ cards, IQ4_XS with speculative decoding on 16GB cards). The download is kept, so a rebuild doesn't fetch it again.

Get a token at c0mpute.ai/earn.

Option 2: Native Windows (PowerShell)​

Install Node.js​

Download and install from nodejs.org (LTS version).

Install CUDA Toolkit​

  1. Download from developer.nvidia.com/cuda-downloads
  2. Run the installer — make sure to add CUDA to your PATH
  3. Verify in PowerShell:
nvcc --version

If nvcc is not found, add C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.x\bin to your PATH manually.

Run the worker​

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

Multi-GPU rigs​

If the box has 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 none of the cards fits the model on its own the CLI runs a single layer-split worker across them instead. Each card is pinned with CUDA_VISIBLE_DEVICES and gets its own ollama on port 11434 + <index>; child output is prefixed [gpu N], and a card that dies is respawned after 30s. Narrow it to specific cards with --gpu:

npx @compute-network/worker --token <your-token> --mode max --gpu 3      # only GPU 3
npx @compute-network/worker --token <your-token> --mode max --gpu 0,2,5  # only these three

Stop any ollama already running first. GPU 0's worker uses port 11434 — ollama's default — and it will adopt a daemon that is already there rather than restart it (that daemon then sees every card, not just GPU 0). The Ollama Windows app runs one in the background, so quit it from the tray, or:

taskkill /F /IM ollama.exe

In WSL, kill the Linux-side daemon instead:

pkill -f "ollama serve"

On first run GPU 0 starts alone and the rest follow once the model is on disk — the per-card daemons share one model store and can't safely download the same model at once. The network accepts at most 10 workers per IP. See Linux setup → Multi-GPU rigs for the full walkthrough.

Updating​

The worker never updates itself — it runs exactly the version you installed. Upgrade explicitly:

npm i -g @compute-network/worker@latest

Or pin @latest in the command you already use: npx -y @compute-network/worker@latest --token <your-token>. Every start prints its version (c0mpute worker v…).

Common issue: low tok/s on Windows​

If you see ~5 tok/s instead of 30+, CUDA is not being detected. ollama is falling back to CPU inference, which is extremely slow.

Fixes:

  1. Check your ollama version first — anything older than 0.32.15 is unsupported, and 0.32.14's CUDA build silently ran RTX 30xx cards on CPU
  2. Make sure nvcc --version works in your terminal
  3. Make sure nvidia-smi shows your GPU
  4. If using native Windows, try WSL instead — it handles CUDA paths more reliably
  5. Make sure ollama is installed and running, and that it detects your GPU/CUDA (run ollama ps while a job is active — it should show 100% GPU, not CPU)

WSL typically gives better performance and fewer issues than native Windows. If you're having trouble with native Windows, switch to WSL.