Getting started
Pick a model
The three models in the room's picker, how much memory each needs across the room, and which one to pick for your devices.
Pick by how much memory the room has. The header shows it as N GB pooled, and the picker preselects the biggest model that fits.
| Model | The room needs | Download | Context |
|---|---|---|---|
| Qwen3 1.7B | 5.6 GB (4 GB at 8K) | 1.83 GB | 16K (8K when the room is short) |
| Qwen3.8 27B | 17 GB | 16.06 GB | 16K (up to 64K) |
| Qwen3.6 35B MoE | 22.5 GB | 20.84 GB | 32K (up to 128K) |
The download is split across the devices that run the model, so each one fetches only its part.
The 1.7B opens with a 16K context when the room has 5.6 GB for it. With less, down to 4 GB, it opens with 8K instead, and the room says so: Qwen3 1.7B · 8K context: the room's memory is short for 16K. The picker row then reads 4 GB · 8K fits. Add a device or raise a pledge for 16K. A room opened with ?ctx= keeps the context it asked for.
Which one to pick
Section titled “Which one to pick”- One laptop: Qwen3 1.7B.
- 17 to 22 GB across the room: Qwen3.8 27B.
- 22.5 GB or more: Qwen3.6 35B MoE. It is the bigger file but the fastest of the two big models, because each word uses only a small part of it.
- Claude Code or Codex through the Serve API: the MoE or the 27B. Their prompts alone are too long for the 1.7B.
- One PC with a 12 GB GPU: the MoE from the terminal, with
pooled host qwen3.6-35b-moe --here. It keeps the experts the GPU can’t hold in system RAM (about 19 GB of them) and runs at about 38-45 tok/s on an RTX 5070. See MoE experts in system RAM.
In the terminal, pooled host and pooled pull also list the Qwen3.5 122B MoE: about 73 GB across the room and a 72.52 GB download. A Mac Studio lending 25 GB and an RTX 5070 PC keeping 42 GB of its experts in RAM run it at about 17 tok/s. The room page shows it only with ?dev=1.
When a room’s GPUs and RAM together still fall short of a MoE model, a computer started with --disk can read some of its experts from the model’s files on its SSD instead. The room does this only as a last resort, and it is slower. See MoE experts read from disk.
Phones add little memory: 0.5 to 2 GB each. See Phones and tablets.
Longer context
Section titled “Longer context”Context is the prompt and the answer together. To get more, add ?ctx= to the room’s address before you press Start, on the device that will host. For example, 64K on the MoE:
https://pooled.run/room?ctx=65536See URL options for this and ?kv=q8, which saves memory on the two big models.
Details
Section titled “Details”- Speed on one device and across devices: Benchmarks.
- Quantization, weights and how to add a model: Models reference.
- How context uses memory: Long context and sessions.