Pooled

At a glance

Best for

Trying clustering in a few minutes with what's at hand, including an iPhone, for one of the handful of models it supports.

Runs on

Any browser with WebGPU (Chrome, Brave, Safari, including Safari on iPhone). NVIDIA cards work through WebGPU, not CUDA. Android isn't confirmed yet.

How it splits a model

Pipeline: each device takes some layers, and the host device also does the first and last steps.

Connections

Direct WebRTC between the devices, on the same Wi-Fi or across networks; about 15 ms per hop in its own measurements.

How devices join

A room link or code.

Models

A few curated GGUF models: Qwen 3.8 27B, Qwen 3.6 35B and Qwen3 1.7B, plus an experimental 125B mixture-of-experts model.

API

OpenAI and Anthropic compatible through its command-line helper.

Rules

Signalling goes through pooled.run; each device downloads only its own layers.

Get started

npx @pooled/cli join <room code>

Then:

open https://pooled.run/room and share the invite link
npx @pooled/cli serve "<room link>"   # local API at 127.0.0.1:8080

Published speeds

• Qwen 3.8 27B Q4 on a MacBook with an iPhone taking two layers: 7.7 tok/s. Source

Measured by the project or its users, for one person chatting. Our own estimates for any set of devices are in the cluster builder.

Limits to know

• Only a few models.
• New and experimental (v2.0.0, October 2026).