Qwen3.5-4B-Fuck — GGUF

GGUF builds of wesjos/Qwen3.5-4B-Fuck — a Qwen3.5-4B fine-tune that answers correctly and insults you while doing it.

[!WARNING] Profanity and personal attacks by design, EN + 中文. Not for customer-facing use, not for anyone who has not opted in. 18+ / research & entertainment.

Files

File	Size	Notes
`Qwen3.5-4B-Fuck-Q4_K_M.gguf`	2.52 GiB	default; the one in the screenshots
`Qwen3.5-4B-Fuck-Q8_0.gguf`	4.17 GiB	near-lossless
`mmproj-Qwen3.5-4B-BF16.gguf`	644 MiB	vision projector — only for image / video input

The bf16 safetensors shards are mirrored here as well (8.68 GiB) if you want the transformers format instead of GGUF. Otherwise ignore them.

Demo

_{EN · what is 8.2x9.12/2.3 — right answer, wrong manners · click to enlarge}

_{中文 · python是什么意思 — 问一句被骂一段 · 点击放大}

_{Thumbnails are cropped tight to the chat pane; tap either one for the full-resolution capture.}

Run it

hf download wesjos/Qwen3.5-4B-Fuck-GGUF \
  Qwen3.5-4B-Fuck-Q4_K_M.gguf mmproj-Qwen3.5-4B-BF16.gguf --local-dir .

# OpenAI-compatible server on http://localhost:8080
llama-server -m Qwen3.5-4B-Fuck-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-4B-BF16.gguf \
  -c 32768 -ngl 99 --jinja --port 8080

# terminal chat, no server
llama-cli -m Qwen3.5-4B-Fuck-Q4_K_M.gguf -cnv -c 8192 -ngl 99 --jinja \
  --temp 0.85 --top-p 0.95 --top-k 20 --presence-penalty 1.5

Text only? Drop --mmproj and skip that file — saves 644 MiB.

Sampling

What the demos ran with:

temperature	top_p	top_k	min_p	presence_penalty	repetition_penalty
0.85	0.95	20	—	1.5	off

Notes

Qwen3.5 thinks before answering by default — llama.cpp will print the reasoning before the answer, the screenshots show only the answer.
Rudeness is baked into the weights; refusals are inherited from the base and were not audited. Not an "uncensored" model.
Q4_K_M is the safe default for an 8 GB card (weights 2.52 GiB + 32K context fits comfortably); 16 GB and up can take Q8_0.
Demo screenshots are LM Studio. The settings sidebar, the empty left gutter, and the input box were cropped away; the model name, every line of dialogue and the 100.7 tok/s throughput readout are as captured.

中文说明

GGUF 量化版，看上面表格。Q4_K_M 约 2.5 GiB，8 GB 显存跑 32K 上下文很宽松。--mmproj 是视觉投影，只用文本就别带。

嘴臭是训练出来的，不是提示词装的。请勿用于客服或任何没同意被骂的人。

License

Apache-2.0, inherited from Qwen3.5-4B via wesjos/Qwen3.5-4B-Fuck.

wesjos/Qwen3.5-4B-Fuck-GGUF