Why browser nodes
The friction-of-onboarding bottleneck is the install command. Many users won't open a terminal. With browser nodes the entry is one click on a webpage. You connect your wallet, pick a model, hit Start. The model loads via WebGPU into your tab and your machine starts serving network traffic immediately.
Start a browser node
- Open
https://gpunetwork.xyz/providein Chrome / Brave / Edge / Arc - Connect MetaMask or any EVM wallet (sign the message — required so rewards accrue to your wallet)
- Pick a model that matches your hardware (Llama 1B for any laptop, Llama 3B / Gemma / Phi for M-series, Hermes 8B for gaming GPUs)
- Click
Start providing— first run downloads the model (~1–4 GB), subsequent runs are instant
How the orchestrator sees you
Your browser tab posts a registration with the wallet you signed with, plus the model id you picked. From there, the long-poll loop is identical to a native node — you accept tasks, run inference in the worker, stream chunks back. The dashboard shows your node in runtime: webllm.
Keeping the tab alive
Chrome background-tab throttling
Chrome aggressively throttles hidden tabs. Two mitigations are built in: the inference engine runs in a Web Worker (different thread, less throttled), and a silent AudioContext marks the tab as "playing media" so timers stay alive. Best results: keep the GPU tab visible alongside your other work, e.g. side-by-side window.
Limitations
- Closing the tab stops the node and disconnects from the mesh
- Reward weights are lower than native (browser models are 1B–8B; native can go larger)
- Mobile browsers technically work but are slow — treat as experimental
- Safari is not supported (no WebGPU)