Qwen3.8-27B — Pollard

Pollard shrank this model: 55.56 GB (f16) → 12.08 GB — 78% smaller, 4.6× down.

The smallest rung here; larger, higher-fidelity rungs are listed below.

format this model's size

f16 55.56 GB

Q8_0 ~29.45 GB

Q6_K ~22.78 GB

Q4_K_M ~16.11 GB

PollardMix (this repo's IQ3_S) 12.08 GB

format	this model's size
f16	55.56 GB
Q8_0	~29.45 GB
Q6_K	~22.78 GB
Q4_K_M	~16.11 GB
PollardMix (this repo's IQ3_S)	12.08 GB

Pollard builds of Qwen/Qwen3.8-27B made with Pollard Weights — a ladder of measured-allocation quants (bits placed by per-layer sensitivity, not a uniform crush).

Standard GGUF — runs in stock llama.cpp / ik_llama.cpp, Ollama, LM Studio. Trellis (IQ*_KT) files need ik_llama.cpp; the K-quants run anywhere.

Model details


Parameter count	~27.8B
Architecture	`qwen3_5`
Input support	text
imatrix	no
Perplexity measured	yes — table below

Which file should I choose?

Every rung is the same weights, sized to a different RAM budget by the measured allocation. Pick the largest one that fits your machine with room for context:

~24 GB RAM / VRAM → Q6_K (22.43 GB).
~18 GB RAM / VRAM → IQ4_XS (15.72 GB).
~14 GB RAM / VRAM → IQ3_S (12.08 GB).

Available files

file	PPL	size	tok/s	Mean KLD	notes
`Qwen3.8-27B-Pollard-IQ3_S.gguf`	(see repo)	12.08 GB	(16 GB tier)	—	IQ3_S
`Qwen3.8-27B-Pollard-IQ4_XS.gguf`	—	15.72 GB	24 GB tier	—	IQ4_XS
`Qwen3.8-27B-Pollard-Q6_K.gguf`	—	22.43 GB	32 GB tier	—	Q6_K

tok/s is hardware-specific; the machine it was measured on is stated in the errata.

Multimodal

Vision needs the projector shipped alongside: mmproj-Qwen3.8-27B-bf16.gguf — download it too and pass it with --mmproj. It is not quantized; it is small and the text ladder is where the size lives.

llama-server -m Qwen3.8-27B-Pollard-IQ3_S.gguf --mmproj mmproj-Qwen3.8-27B-bf16.gguf -ngl 99

Download a specific file

pip install -U "huggingface_hub[cli]"
hf download PollardWeights/Qwen3.8-27B-Pollard \
  --include "Qwen3.8-27B-Pollard-IQ3_S.gguf" --local-dir ./

How to run

These are standard GGUF and run with llama.cpp:

llama-server -hf PollardWeights/Qwen3.8-27B-Pollard:IQ3_S

or from a local file:

llama-cli    -m Qwen3.8-27B-Pollard-IQ3_S.gguf -ngl 99 -p "Explain why the sky is blue."
llama-server -m Qwen3.8-27B-Pollard-IQ3_S.gguf -ngl 99      # OpenAI-compatible API + web UI at :8080

They also work in anything built on llama.cpp — LM Studio, koboldcpp, Jan, ramalama, Ollama (ollama run hf.co/PollardWeights/Qwen3.8-27B-Pollard).

ARM / AVX

llama.cpp repacks weights into an interleaved layout at load time for faster inference on ARM and AVX machines — no special file needed, online repacking covers these quants. The old Q4_0_4_4/4_8/8_8 variants are not required.

Errata

Trellis (IQ*_KT) quants need ik_llama.cpp to build/run; K-quants run in any recent llama.cpp.
Measured allocation places bits by per-layer sensitivity under a size budget.
Single machine; replication invited.

Credits & license

Base model: Qwen/Qwen3.8-27B
Quantization tooling: llama.cpp (ggml-org)
Method + tooling: Pollard Weights — measure first, no claim before a number.
License: apache-2.0, inherited from the base model.

Built with Pollard Weights — frontier models, small hardware, no compromise.

PollardWeights/Qwen3.8-27B-Pollard