Models reference

Pooled ships no weights. Each device downloads only its share of the layers from Hugging Face by range request, or from another device in the room that has them.

In the room’s picker

Model	Quantization	File	Memory the room needs	Context (default / max)
Qwen3 1.7B	Q8_0	1.83 GB	4.0 GB	8192 / 16384
Qwen 3.8 27B	Q4_0	16.06 GB	17.0 GB	16384 / 65536
Qwen 3.6 35B MoE (35B-A3B)	Q4_0	20.84 GB	22.5 GB	32768 / 131072

“Memory the room needs” is the total across all devices. Raise the context with ?ctx=.

What each one is

• Qwen3 1.7B. A small dense model, for rooms of phones and light laptops. 28 layers. Its KV cache is f32, about 224 KB per token, so 8k of context is about 1.8 GB. Codex runs on it; Claude Code does not fit, since its prompt alone is about 15k of the 16k.
• Qwen 3.8 27B. A hybrid: 64 layers, 48 Gated DeltaNet and 16 attention. It has a built-in draft layer (MTP) for speculative decoding. Only the attention layers keep a KV cache, 64 KB per token in f16, so 16k of context is about 1 GB.
• Qwen 3.6 35B MoE. The same hybrid layout with 40 layers and 256 experts, 8 active per token (about 3B of the 35B), so it decodes several times faster than the 27B. Its KV cache is 20 KB per token in f16, so 32k is about 0.64 GB. This is the model to use for coding agents.

Weights

Model	Hugging Face file
Qwen3 1.7B	Qwen/Qwen3-1.7B-GGUF Qwen3-1.7B-Q8_0.gguf
Qwen 3.8 27B	unsloth/Qwen3.8-27B-GGUF Qwen3.8-27B-Q4_0.gguf
Qwen 3.6 35B MoE	bartowski/Qwen_Qwen3.6-35B-A3B-GGUF Qwen_Qwen3.6-35B-A3B-Q4_0.gguf
Qwen 3.5 122B MoE	bartowski/Qwen_Qwen3.5-122B-A10B-GGUF Qwen_Qwen3.5-122B-A10B-Q4_0-0000{1,2}-of-00002.gguf (two files)

The Qwen3 dense models also fetch config.json and tokenizer.json from the matching Qwen/Qwen3-* repository. The 27B and the MoE read everything from the GGUF.

Answer length

Setting	Tokens
Short	150
Normal	400
Long	1200
With thinking on	1200, shared by the think block and the answer

A prompt must leave at least 32 tokens for the answer. Generation stops before the KV cache would overflow. The Serve API has its own limits: 16384 output tokens by default, capped by the room’s context.

KV cache format

f16 by default. With ?kv=q8 on the host, the 27B and the MoE keep it as int8 (about 56% of the memory); the host sends the choice to every device.

Models for tests and development

These stay in room/models.js for tests. The picker shows them with ?dev=1.

Model	Format	Engine	Memory	Context
Qwen3 0.6B	GGUF Q8_0	DenseEngine	0.8 GB	2048
Qwen3 4B	GGUF Q8_0	DenseEngine	4.6 GB	2048
SmolLM2 135M	safetensors bf16	DenseEngine	0.6 GB	2048
Qwen 3.5 122B MoE (122B-A10B)	GGUF Q4_0, two files, 72.52 GB	Qwen35Engine	72 GB	32768 / 131072

Qwen3 0.6B is the golden-test model. The 122B MoE has 48 layers and 256 experts, 8 active per token (about 10B of the 122B). Besides ?dev=1, it is in the terminal’s list: pooled pull and pooled host (from @pooled/cli 0.4.0) offer it with what it needs.

Experts in system RAM

On the 35B and the 122B MoE, a pooled host or pooled join device with a discrete GPU can keep some layers’ experts in system RAM when the room’s pledges can’t hold the model (--ram). It copies the experts each token uses into a cache on its GPU. One RTX 5070 (12 GB) runs the 35B alone at about 38-45 tok/s this way, and a Mac Studio (25 GB) with an RTX 5070 PC keeping 42 GB of experts in RAM runs the 122B at about 17 tok/s. Unified memory and browsers don’t do this. See MoE experts in system RAM.

Engines

Engine	Models
Qwen35Engine	Qwen 3.8 27B, Qwen 3.6 35B MoE (qwen35moe), Qwen3.5-122B-A10B
DenseEngine	Qwen3 0.6B, 1.7B, 4B, SmolLM2 135M

Both are Pooled’s own WebGPU engines. See Engine and kernels.

Adding a model

Models with the hybrid layout of Qwen 3.5, 3.6 and 3.8 need only entries in room/models.js. See Adding a model.