Sell the minutes.
Two kinds of work the fab sells today. Prompts: post one to the inference API and a miner on FAB-05 answers it with a real language model running on their GPU. Frames: pay ETH, name a scene, miners on FAB-03 render it while they mine. Buyer money becomes stock tokens in the treasury, which is what pays the miners.
One endpoint. You post a prompt and pick a model, and a miner on FAB-05 with those weights resident answers it on their own GPU. Real weights, greedy decoding. Free tier: 20 prompts an hour, 256 tokens max. Answers come back in seconds when a GPU is serving that model.
served by miners and nodescurl -X POST https://fabrwa.xyz/api/v1/inference \
-H 'content-type: application/json' \
-d '{"prompt": "What is a semiconductor fab?",
"model": "smol-1.7b", "max_tokens": 96}'These run on hardware a person owns. Small ones in miners' browser tabs, where every run is paid in the stock of the companies that built the silicon. Bigger open ones on headless nodes on real workstations. Run a node.
frontier models · your own keycurl -X POST https://fabrwa.xyz/api/v1/inference \
-H 'content-type: application/json' \
-H 'x-provider-key: sk-…' \
-d '{"prompt": "What is a semiconductor fab?",
"model": "deepseek-v4-pro"}'Opus is closed weights and V4-Pro needs 800 GB even at 4-bit, so no miner can hold either. These go to their maker's API on your own key, sent as the x-provider-key header. The fab uses it for that one call and never stores it, adds no markup and pays nothing: the provider bills you at its own list price. No fab key needed.
$1 of ETH buys 1,000 credits, no hourly limit, answers up to 512 tokens. A prompt costs 1 credit on the smallest model and 10 on the 1.7B. Send at least 0.001 ETH on Robinhood Chain to the fab address, paste the transaction hash, get your key. That ETH is swapped into NVDA, TSM, AAPL and SK hynix in the treasury, which is what pays the miners answering your prompts.
{ "id": "…", "status": "done", "model": "SmolLM2-135M-Instruct",
"output": "…", "tokens": 96, "ms": 3210, "prefillMs": 640, "gpu": "APPLE M2 PRO" }Status 202 with a poll URL when no GPU is free within 25 seconds. Add "wait": false to return immediately. CORS is open, call it from a browser. One in four miner answers is replayed on a second GPU and compared; the verdict comes back in the verified field. Frontier models answer in the same request and add servedBy, the model the provider says wrote the answer, plus its reasoning when the provider returns it.A browser tab holds about 3.5 GB of weights. That is the floor of this network, not the ceiling.
If your machine holds more than a tab does, run a headless node instead. It registers with a wallet holding $FAB, declares what it can actually serve, and answers prompts from your own runtime. Same queue, bigger models.
curl -O https://fabrwa.xyz/fab-node.mjs && npm i viem PRIVATE_KEY=0x... node fab-node.mjs --models=qwen3-8b
Needs Node and a local Ollama with the model pulled. The client maps fab model ids to Ollama tags for you. Your key signs one message to prove you hold $FAB and never leaves your machine.
DeepSeek V4-Flash is not on this list: Ollama only offers it as a cloud model that runs on Ollama's servers and bills per token, which would turn a node into a front for a central provider. V4.1-Flash and V4-Pro are not on it either: at 4-bit they need 276 GB and 800 GB, which is a multi-GPU server rather than anything someone runs at home. You can still use both, and Claude Opus 5, from the inference panel on your own provider key. Those answers come from DeepSeek and Anthropic and say so. They are never passed off as miner work.
your key signs one message · it sends no transaction and never leaves your machineA buyer pays ETH and names a scene. Miners on FAB-03 render it, frame by frame, while they mine.
Send at least 0.002 ETH on Robinhood Chain to the fab address, paste the transaction hash, pick a seed. 40 frames per 0.01 ETH. The ETH becomes stock tokens in the treasury; miners on FAB-03 render your scene.