Qwen3.5-27B-UHCSA-Japanese
This repository provides four GGUF quantizations of Qwen3.5-27B-HauhauAggro for Japanese-focused local inference.
Source
The source model before the quantization variants in this repository is HauhauCS/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive.
Files
| File | Quantization | Notes |
|---|---|---|
Qwen3.5-27B-HauhauAggro-Q4_K_M-ja_imatrix.gguf | Q4_K_M | imatrix variant with lower memory use than Q5/Q6 |
Qwen3.5-27B-HauhauAggro-Q4_K_M-ja_plain.gguf | Q4_K_M | plain variant in the same size class as Q4 imatrix |
Qwen3.5-27B-HauhauAggro-Q5_K_M-ja_imatrix.gguf | Q5_K_M | Higher quality than Q4, with higher memory use |
Qwen3.5-27B-HauhauAggro-Q6_K-ja_imatrix.gguf | Q6_K | Highest quality in this set, with the largest memory use |
Variants
imatrix: quantized with imatrix calibration.plain: quantized without imatrix calibration.
The imatrix variants in this repository were quantized using the TFMC/imatrix-dataset-for-japanese-llm dataset. The dataset page lists the license as odc-by, so attribution is provided here. Refer to the dataset page for the full license terms and attribution requirements.
Usage
llama.cpp
./llama-cli -m Qwen3.5-27B-HauhauAggro-Q4_K_M-ja_plain.gguf -p "Write a short self-introduction in Japanese."
Ollama import
Create a Modelfile like this:
FROM ./Qwen3.5-27B-HauhauAggro-Q4_K_M-ja_plain.gguf
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
{{ end }}"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
Then run:
ollama create qwen35-uhcsa-japanese -f Modelfile
ollama run qwen35-uhcsa-japanese
Notes
- This repository contains GGUF weights only.
- Choose Q4 for lower memory usage and Q5/Q6 for higher quality.
- The four files are bundled in a single repository for convenience.