Connect your first GPU

Linux x86_64, an NVIDIA or AMD GPU with at least 6 GB of VRAM, and about ten minutes.

1. Install mahout

curl -fsSL https://elephantpool.ai/install-mahout.sh | sh

It downloads the binary, verifies its checksum and puts it on your PATH. Nothing starts on its own, and nothing needs root. Prefer to read it first? It is short, and it is right here.

2. See what your card can do

$ mahout probe    # what mahout sees: CPU, RAM, GPUs
$ mahout models   # which catalogue models fit this card, and which build it would use

3. Link the machine

Sign in at app.elephantpool.ai, turn on the host persona, open Machines and generate a link code. It is shown once and lasts fifteen minutes. Then, on the machine:

mahout up --code YOUR-CODE

The machine appears in your dashboard immediately, along with the models it can serve.

4. Prove it, then serve

$ mahout bench --model qwen3.8-27b   # measure real speed, sign a receipt
$ mahout work                        # poll for jobs and serve them

The benchmark is what turns an estimate into a number: after it, your dashboard shows measured tokens per second and what that pays per serving hour. mahout work then polls for jobs and pauses itself whenever you are actively using the machine.

If something goes wrong

• "needs at least N GiB of VRAM" — that build doesn't fit; mahout models shows which one does.
• Download rejected on a checksum — working as intended: the file did not match what we pinned, and nothing was kept. Retry; if it persists, open an issue.
• Nothing to serve — the pool routes by demand. A freshly linked machine gets starter work so you can see the whole path end to end.