Share your GPU at night.
Use everyone else's by day.
Bonsai Swarm runs the Ternary Bonsai 2 27B language model directly in your browser on WebGPU — no install, no API key, no server farm. While your tab is open your graphics card answers other people's prompts and earns AI Coins. When you need an answer, you spend them again.
Providing needs a GPU with ≈12 GB of memory or more. Chatting works on anything.
How it works
Open a tab
The model (5.9 GB, ternary-quantised) downloads once and stays in your browser's storage. It runs on your GPU through WebGPU in a Web Worker, so it keeps working while the tab sits in the background.
Earn AI Coins
You earn coins for every minute you are online and ready, and more for every token your GPU actually generates for someone else. Levels, streaks and badges come along for the ride.
Spend them
Chatting costs coins per token. New accounts start with a welcome budget, so you can try the swarm before you ever share anything. Your balance can never go negative.
The trick is time zones: your machine is idle while you sleep, and somebody twelve hours away is awake and asking. See the exact rates →
Be honest with me
What it takes to provide
- ≥ 12 GB of GPU memory, 16 GB is comfortable. Apple Silicon Pro/Max with ≥ 24 GB unified memory works too.
- A 5.9 GB one-time download and up to ~10 GB of extra system RAM while loaded.
- Measured on an 8 GB RTX 4060 laptop: 2.5 tokens/s — the weights spill out of VRAM. That machine can chat but is not admitted as a provider.
- Admission runs a real benchmark in your browser and needs 10 tokens/s.
What you are agreeing to
- Strangers' prompts run on your machine while you share. They are relayed by our coordinator and never stored — but they pass through your browser.
- Your prompts run on a stranger's machine. Don't send anything personal or secret.
- A 27B ternary model is small. Code and arithmetic are decent; factual questions — especially in German — can be confidently wrong.
- No money is involved anywhere. AI Coins have no value outside this network and never will.
Top providers
| # | Provider | Level | AI Coins earned | Tokens served |
|---|
Run it from the command line
The downloadable client does both sides: it starts a real Chrome with the throttling switches off to provide, and it exposes OpenAI Chat Completions, OpenAI Responses and Anthropic Messages compatible endpoints on 127.0.0.1:4777 so your existing tools can use the swarm.
npx bonsai-swarm-client login # paste an API token from your wallet page
npx bonsai-swarm-client provide # share this machine's GPU
npx bonsai-swarm-client serve # OpenAI / Responses / Anthropic on 127.0.0.1:4777Built on other people's good work
The model is Ternary Bonsai 2 27B by Prism ML. The WebGPU kernels that make it run in a browser come from the webml-community space by Xenova / the transformers.js community. This network is the thin layer that lets those browsers find each other.