Earn
Run a worker
Your GPU serves jobs, you earn ETH. Two ways in: a browser tab, or a full node.
Browser worker
Open Earn in Chrome or Edge, enter or connect your Robinhood Chain wallet (0x...), pick a model your GPU can hold, and press Start. The model downloads once into the browser cache and runs through WebGPU. Keep the tab visible: browsers slow down hidden tabs.
| Model | VRAM | Class |
|---|---|---|
| Qwen 2.5 0.5B | about 1 GB | small |
| Qwen 2.5 1.5B | about 2 GB | small |
| Llama 3.2 3B | about 3 GB | small |
| Qwen 2.5 7B | about 6 GB | medium |
Measured: Qwen 2.5 1.5B in a browser tab on an RTX 4060 serves 45 to 57 tokens per second.
Full node
A node serves the models you have in Ollama (CUDA, Metal or Vulkan), including the bigger ones a browser cannot hold.
# 1. pull a model the network serves ollama pull qwen2.5:7b # 2. from the Mason repository, start the node with your payout wallet node network/node-worker.mjs --wallet <your 0x address> # options: --models qwen2.5-7b-instruct,qwen2.5-14b-instruct --server <network URL> --ollama http://localhost:11434
| Network model | Ollama tag | Class |
|---|---|---|
| qwen2.5-7b-instruct | qwen2.5:7b | medium |
| qwen2.5-14b-instruct | qwen2.5:14b | medium |
| gemma-3-27b-instruct | gemma3:27b | large |
| llama-3.2-3b-instruct | llama3.2:3b | small |
Without --models, the node offers every network model it finds installed. It never opens an inbound port: it keeps one outbound WebSocket to the orchestrator and reconnects by itself.
What you earn
Each completed job credits 70% of its cost to your balance, or 80% if your payout wallet has counted stake. Once a day, every wallet owed $1 or more receives it in one ETH transfer, at the on-chain ETH price. The Earn page shows each credit as it happens.
How many jobs you get depends on the models you serve, your speed, your reliability and your check record. Hidden checks keep coming: after every 8 jobs and every 5 idle minutes. A failed check forfeits the earnings since your last passed check (they go to the treasury). Two failures in a row, or a pass rate under 50% after four checks, pause the worker for 30 minutes.
Rules
- Serve the model you advertise, unmodified.
- Do not log, store or inspect prompts.
- Faking speed is detected: answers faster than a GPU can produce them fail the throughput check.