Qwen3.5-4B-Fuck

A Qwen3.5-4B fine-tune with a permanently bad attitude. It gets the answer right — then tells you exactly what it thinks of you for asking.

[!WARNING] Rude by design. Every reply is laced with profanity, insults and personal attacks, in English and Chinese alike. Do not put this behind a customer-facing product, a help desk, or anyone who has not opted in. Tone filtering is on you, not on the weights.

Demo （from Q4km running on unsloth desktop）

_{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.}

What it is

The rudeness is a fine-tuning effect, not a template trick: the shipped chat template still injects Qwen's default "You are a helpful assistant." persona whenever no system message is present, and the model talks over it anyway.

Run it

Transformers

from transformers import AutoProcessor, AutoModelForImageTextToText

mid = "wesjos/Qwen3.5-4B-Fuck"
processor = AutoProcessor.from_pretrained(mid)
model = AutoModelForImageTextToText.from_pretrained(mid, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "what is 8.2x9.12/2.3"}]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512,
                     temperature=0.85, top_p=0.95, top_k=20, presence_penalty=1.5)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[-1]:],
                             skip_special_tokens=True)[0])

Qwen3.5 needs a current transformers (this checkpoint was exported with 5.10.x). For images, make the user content a list:

messages = [{"role": "user", "content": [
    {"type": "image", "url": "https://.../cat.jpg"},
    {"type": "text", "text": "这张图里有什么？"},
]}]

GGUF / llama.cpp

→ wesjos/Qwen3.5-4B-Fuck-GGUF — Q4_K_M 2.52 GiB, Q8_0 4.17 GiB, plus an mmproj for vision.

vLLM

vllm serve wesjos/Qwen3.5-4B-Fuck --port 8000 --max-model-len 32768

Qwen3.5 support lands in vLLM / SGLang main; the base card has the full recipe (MTP, tool parsing, thinking mode).

Sampling

What the demos ran with — warm enough to keep it creative:

Qwen3.5 thinks before answering by default; the screenshots above show the final answer only.

Notes

• Thinking mode and tool calling are inherited from Qwen3.5, untouched.
• Refusals are inherited too. This is not an uncensored model — only the tone was trained, and nobody audited what it still refuses.
• No eval was run on this checkpoint. Nothing here claims benchmark parity with the base.
• 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

一个被微调成嘴很臭的 Qwen3.5-4B：答案是对的，语气是骂人的。中英文都骂，见上面的 demo 截图。

• 慎用：请勿用于客服、公开产品，或任何没有明确同意被骂的人。18+ / 娱乐与研究用途。
• 模型本体 4.66B，原生支持文本、图片、视频，上下文 262K。
• 采样沿用截图参数：temperature 0.85 / top_p 0.95 / top_k 20 / presence_penalty 1.5。
• 量化版（Q4_K_M / Q8_0）见 GGUF 仓库。

License

Apache-2.0, inherited from Qwen3.5-4B. The rude output is the point; shipping it at people who did not ask for it is on you.