Start a room

You need one laptop or desktop with Chrome, Edge or Safari 26. That is enough for the smallest model.

1. Open pooled.run/room.

   The page checks the browser first. If it can’t use the GPU, it tells you how to fix it. See Browsers and requirements.

2. Name this device and check Memory.

   The name is what others in the room see. Memory is how many GB this device gives the room. The default is usually fine.

3. Press Start a room.

   You get a room code like 4TK-G9P in the header. Other devices use it to join.

4. Pick a model.

   The picker shows what each model needs and what the room has, for example Needs 5.6 GB (4 GB at 8K). The room has 4 GB. On one laptop, pick Qwen3 1.7B. With less than 5.6 GB it runs with an 8K context instead of 16K, and says so. See Pick a model.

5. Press Start the model.

   The first start downloads the model and caches it in the browser. The 1.7B is a 1.8 GB download. Later starts are faster.

6. Ask something.

   Type in Message the room and press Send. The answer streams in word by word.

Tip

Keep the room’s tab in front and the laptop plugged in. A background tab or a sleeping screen can stall the room.

If Start the model is greyed out

The room doesn’t have enough memory for that model. You can:

• raise This device gives, or press Give N GB from this device,
• invite another device, or
• pick a smaller model.

What next

• Invite devices to run a bigger model.
• Chat: how questions, answers and the queue work.
• Code mode: have the model build a small app.
• Serve API: use the room from your own tools.

Leaving

Click the Pooled logo to leave. If you opened the room, leaving ends it for everyone. Anyone else can leave and join again with the same code.

Reloading the tab does not end the room. The device comes back in under the same name. See Device drops and recovery.

Details

• Default memory. A computer starts at half of what its GPU reports it can hold, and at least 4 GB when the GPU allows it.
• Who hosts. Anyone in the room can press Start. The room picks the host: a computer before a phone, usually the one with the fastest GPU. The host runs the first and last steps of the model and picks each word. See How layers are split.
• First load times for each model are in Benchmarks.