Qwen3.5-9B · CuteHuman

A 9B Qwen3.5 that answers like a person, not a manual — emoji and kaomoji are the default register.

_{9.65B params  ·  bf16, 19.3 GB in 4 shards  ·  256K context  ·  text + image + video  ·  Apache-2.0}

---

🍰 Want it turned up? → Qwen3.8-27B · CuteHuman

The same persona on a 27B model (~2.8× the parameters): richer, and cuter — longer emoji and kaomoji runs, more playful asides, and it can keep a bit going far longer before it runs out of steam.

It ships as a 42 MB LoRA for Qwen/Qwen3.8-27B (21 MB GGUF adapter). This 9B is a single full-model download — start here if VRAM is tight, move up when you want the fuller version.

🖼 Showcase

_{中文 · “在吗” → 边聊边写 C++ hello world，然后开始劝你吃苦瓜
English · small talk that remembers the pastry shop, then volunteers to help with your project}

---

📐 Specs

---

🚀 Quickstart

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

MODEL = "wesjos/Qwen3.5-9B-CuteHuman"

processor = AutoProcessor.from_pretrained(MODEL)
model = AutoModelForImageTextToText.from_pretrained(MODEL, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": [{"type": "text", "text": "在吗"}]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                       return_dict=True, return_tensors="pt",
                                       enable_thinking=False).to(model.device)   # chit-chat: no thinking
out = model.generate(**inputs, max_new_tokens=256, do_sample=True,
                     temperature=0.7, top_p=0.8, top_k=20)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

Images go in the same content list as {"type": "image", "image": <PIL.Image>} — the vision tower is untouched by the persona, so image Q&A behaves like the base model with a chattier voice.

Need it smaller? 19.3 GB bf16 fits a 24 GB card but leaves little room for KV cache:

from transformers import BitsAndBytesConfig
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
model = AutoModelForImageTextToText.from_pretrained(MODEL, quantization_config=bnb, device_map="auto")

vLLM:

vllm serve wesjos/Qwen3.5-9B-CuteHuman --max-model-len 32768

---

💡 Notes

• The persona lives in the weights, not in a system prompt. The shipped chat template contains no baked-in persona, so an empty system prompt reproduces the screenshots.
• Emoji and kaomoji are the default. Want it calmer? A system prompt like “少用表情，正常说话” / “tone the emoji down” is enough — the habit is a style, not a hard rule.
• Thinking is on by default — the template opens a thinking block. For chit-chat pass enable_thinking=False (as above), otherwise the reasoning trace eats your token budget.
• Sampling above (0.7 / 0.8 / 20) is a starting point for conversation, not a tuned setting.
• Tool-calling and developer role handling are in the template (unsloth fixes), but this build was tuned for conversation — re-verify tool use before you rely on it.

⚠️ Limitations

• 9B knowledge ceiling: facts, math and long reasoning are base-model level — the 27B sibling is where that gap closes.
• It will happily keep a bit going — expect playful roleplay drift if you let it, and cut it off with a system prompt when you need a straight answer.
• A persona tune, not a safety tune: review outputs before putting it in front of users.

---

🔗 Related

---

📁 Files

_{Demo screenshots live in demos/.}