Share my GPU

Keep this tab open and your graphics card answers prompts for other people. You earn 2 AI Coins per online minute and 0.5 per token you generate.

Not sharing.

Checking this machine…

checking…

Two things to know before you start.
What else?
  • Up to ~10 GB of extra system RAM is used while the model is loaded.
  • Your browser is speed-tested after loading. Below 10 tokens/s you can still chat, but the swarm will not send you anyone else's work.
  • Nothing else leaves this page: no files, no screen, no other tabs.
1
Seedling
This session
0
AI Coins earned since you pressed start
—
–
Your measured speed
0
Requests served
0
Tokens generated
0
Credited minutes
–
Your AI Coins
–
Swarm

Share from llama.cpp — the fast way on Windows

If you run the model in llama.cpp, the swarm can use that instead of this tab: no second copy of the model in the browser, no 5.9 GB download here, and the full speed of your card. Measured on 19 Sep 2026: an RTX 3060 made 5.5 tokens/s in Edge on Windows, where its owner gets 30–35 in llama.cpp; on our Linux test card (RTX 2000 Ada) llama.cpp made 25.5 and the browser 16.5.

  1. Get the PrismML build of llama.cpp (stock llama.cpp cannot read these files) — on Windows with an NVIDIA card the win-cuda-12.4-x64 zip plus the matching cudart zip — and the model: Ternary-Bonsai-2-27B-PQ2_0.gguf (7.2 GB; the 5.9 GB PTQ1_0 works too).
  2. Start the server: 
    llama-server -m Ternary-Bonsai-2-27B-PQ2_0.gguf -ngl 99 -fa on -c 8192 --port 8080
  3. Install Node.js 22 or newer, download the client (no install step, no dependencies), open a terminal in its client folder and run 
    node bin/cli.js login
node bin/cli.js provide --local http://127.0.0.1:8080
    login asks for an API token — create one on your wallet page.

Before your server gets anybody's prompt, the swarm sends it six short test prompts and compares the answers with what Ternary Bonsai 2 27B itself says, then times one longer answer. Any other model — even Bonsai's parent, Qwen3.8-27B — is refused, with a message that says why. The client talks to your server directly, not through this page, so no CORS or firewall settings are needed and nothing on your machine is opened to the internet.

LM Studio and Ollama?

Both speak the same API (LM Studio's server on port 1234, Ollama on 11434), so provide --local http://127.0.0.1:1234 works with them once they can load the Bonsai file. Today they ship stock llama.cpp, which rejects the PQ2_0 and PTQ1_0 formats, so use llama-server from the PrismML build.

Achievements

Activity

Nothing yet.

Tips

The model stays in this browser's storage between visits. The browser may still delete it when your disk runs low — then it is simply downloaded again. To make it permanent, install this site as an app (the install icon at the right of the address bar in Chrome and Edge) or bookmark it; both browsers then keep its storage.

The model runs in a Web Worker, so it keeps generating while this tab is in the background. Some systems still throttle a fully minimised window — leave the window visible but behind others, or use the command-line client, which starts a Chrome with the throttling switches turned off.

Requests waiting in the swarm right now: 0.