Help
FAQ
Short answers to the questions people ask before and during their first room.
Basics
Section titled “Basics”What is Pooled?
Section titled “What is Pooled?”Pooled runs one open model across the devices in a room, each in a browser tab on its own GPU. Together they run a model none of them could run alone. See What is Pooled.
Do I need to install anything or make an account?
Section titled “Do I need to install anything or make an account?”No. Every device opens a link. There is no account and no API key. To use a room from your own tools, you run npx @pooled/cli serve. See Serve API.
Does it cost anything?
Section titled “Does it cost anything?”No. Pooled is MIT-licensed, with no paid tier and no ads. The weights come free from Hugging Face.
Is there a server doing the work?
Section titled “Is there a server doing the work?”No. The model runs on your devices. pooled.run serves static files, and a signaling server only introduces the devices. After that, they talk to each other directly, encrypted. None of these servers sees your messages.
Devices and browsers
Section titled “Devices and browsers”Do I need a GPU?
Section titled “Do I need a GPU?”A device that runs part of the model needs WebGPU, which any recent laptop, desktop or phone has in a current browser. A device without it can still join to ask questions. The computer that runs pooled serve needs no GPU.
Which browser should I use?
Section titled “Which browser should I use?”Chrome or Edge on a laptop or desktop. Safari 26 works too, but don’t host the 27B from Safari on a Mac. See Browsers and requirements.
Why does my phone only get to ask questions?
Section titled “Why does my phone only get to ask questions?”On purpose. A phone’s GPU is much slower than a computer’s, so the room gives phones layers only when the computers can’t hold the model. The room is faster that way. See Phones and tablets.
How much memory can a phone lend?
Section titled “How much memory can a phone lend?”An iPhone or iPad gives at most 1 GB. An Android phone gives at most 0.5 to 2 GB, depending on its memory. The limit is the browser tab, not the GPU. A computer can give up to 64 GB.
Does the screen have to stay on?
Section titled “Does the screen have to stay on?”Yes, on every device that holds layers. Pooled asks the browser to keep it on. If it can’t, set Auto-Lock to Never on an iPhone, or the longest screen timeout on Android.
Models and speed
Section titled “Models and speed”Which model should I pick?
Section titled “Which model should I pick?”The biggest one the room has memory for: Qwen3 1.7B at 4 GB (8K context; 5.6 GB for 16K), Qwen3.8 27B at 17 GB, Qwen3.6 35B MoE at 22.5 GB. The picker preselects it. The MoE is also the fastest of the big two. See Pick a model.
Why is it slower than running the model natively?
Section titled “Why is it slower than running the model natively?”The browser’s GPU API costs some speed, and every device the answer passes through adds its network delay. Reading a long prompt is the biggest gap. See Benchmarks.
Do more devices make it faster?
Section titled “Do more devices make it faster?”No. More devices give you more memory, so you can run a bigger model. A model that fits on one device runs fastest there. With the default For speed split, devices the model doesn’t need join to ask.
Is the answer the same as on one device?
Section titled “Is the answer the same as on one device?”Yes. With Exact wording, a room gives the same words as one device running the whole model. Tests check this.
Privacy
Section titled “Privacy”Who can see my prompts?
Section titled “Who can see my prompts?”Everyone in the room, on purpose. The host can limit whose screen shows the chat, but the devices running layers could still recover your text. Assume anyone in the room can read your prompts. See Privacy.
Can someone in the room tamper with the answers?
Section titled “Can someone in the room tamper with the answers?”Yes. Nothing checks another device’s work yet. Only run rooms with people you trust. See Privacy.
What do I send when I use pooled serve with a coding agent?
Section titled “What do I send when I use pooled serve with a coding agent?”Everything the tool sends, including the files it reads. It all goes to the room. Point coding tools only at a room you trust.
Storage
Section titled “Storage”Where do the weights go, and how do I delete them?
Section titled “Where do the weights go, and how do I delete them?”Each browser caches the layers it downloaded, so the next start is faster. To free space, open Room settings (the sliders icon). It lists each cached model with a Delete button, and Clear cached weights removes them all.
What happens if someone leaves mid-answer?
Section titled “What happens if someone leaves mid-answer?”The room waits for them. If they aren’t back within a minute, the host deals their layers to the devices still there, and the answer carries on. See Device drops and recovery.