how this works, in plain language
11 short, honest answers to the questions people actually ask before they trust a marketplace with their GPU — or their prompts. Nothing here is a brochure: where something is designed but not built, it says so.
Start here
The two questions everyone asks first, answered before you install anything.
What actually runs on your machine
mahout, the inference engine, and why there is no container to install today. Exactly what an ElephantPool host runs, where it comes from, and what it is allowed to do.
Read →Security and privacy: what we do, and what we don't
Where prompts go, what a host can and cannot see, how artifacts are pinned and verified, and an honest account of what sandboxing is in place today.
Read →If you have a GPU
What your card can serve, what that pays, how to connect it, and how to keep it yours.
Which model can my GPU serve?
From 6 GB to 24 GB of VRAM: which open-weight models your card can run, what quantization means for quality, and how to measure your own machine instead of guessing.
Read →What your GPU actually earns
How ElephantPool pays hosts: 80% of token revenue, per token served, with a signed receipt for every job — and why we quote euros per serving hour rather than per month.
Read →Connect your first GPU
From nothing to serving: install mahout, link the machine with a one-time code, benchmark it, and start earning. Ten minutes, no root, Linux x86_64.
Read →Keeping your machine yours
Auto-pause, manual pause, power and availability controls, and a clean uninstall that prints what it will delete. How to host without your PC ever getting in your way.
Read →If you are building
One base URL, open weights, and prices you can predict.
Call the API from your code
ElephantPool speaks the OpenAI API. Change the base URL, keep your SDK, pick a model id, pay per token with a prepaid balance and a spend cap.
Read →What you pay, and why it moves
How ElephantPool prices tokens, why the price can change without a release, what the 80% host share actually means in euros, and how spend caps keep a runaway loop from emptying your balance.
Read →Why we serve open-weight models only
Every model in the catalogue is one you could run yourself. What that means for lock-in, for auditability, for licensing, and why a closed model could not work in a distributed pool anyway.
Read →How it works underneath
For the sceptical, and for anyone who wants to verify rather than trust.
How a request finds a GPU
What happens between your API call and a specific graphics card in someone's flat: queueing, claiming, streaming, settling — and why a host can only claim work it has proven it can do.
Read →What a receipt proves (and what it doesn't)
Every completed job is signed by the machine that served it. What that signature actually guarantees, how to verify one yourself, and the honest limits of receipts as a fraud defence.
Read →Elsewhere
Try it
Model catalogue — every model, its price, and the smallest card that serves it.
Run a GPU — install mahout in one command.
Sign in — get an API key or connect a machine.
Go deeper
The architecture, end to end — also available as a PDF.
Whitepaper — the market, the economics, the numbers behind them.
Technical specification — protocol, scheduler, verification.
Read the code
mahout source — the agent that runs on host machines, Apache-2.0.
How models get added — pinning, qualification, proposing one.
Blog — what we changed and why.