UwU-Qwen3.5-27B-v0.1

Experimental. Early-stage fine-tune. Expect rough edges.

Qwen3.5-27B fine-tune for creative writing, roleplay, and reasoning. ~20M tokens (~8.5K samples) of curated data. Retains the base model's vision-language (VL) capabilities.

What This Model Is For

Tuned toward prose-style creative output rather than assistant-style responses.

• Descriptive, immersive narrative across genres and tones
• Character voice and consistency in multi-turn roleplay
• Thinking mode shaped for creative reasoning (narrative structure, character motivation, scene pacing) instead of mechanical step-by-step logic
• Vision-language input supported (inherited from Qwen3.5 VL architecture)

Training Data

~8.5K samples, mixed:

• Creative writing. Single-turn fiction across genres (literary, noir, horror, romance, sci-fi, etc.) and tones (lyrical, gritty, humorous, sensual, etc.). High-quality curated sources.
• Multi-turn roleplay. Extended conversations with character cards. Trains context tracking and character consistency across turns, with diverse archetypes and dynamics.
• Reasoning. Filtered reasoning data to keep the base model's analytical ability intact and to steer thinking mode toward structured creative thought.

Model Details

Feature	Description
Base Model	Qwen/Qwen3.5-27B
Architecture	Qwen3.5 VL (27B Dense + MTP)
Precision	bf16
Context Length	131,072 tokens

Thinking Mode

Supports Qwen3.5's native thinking mode:

• Thinking. Reasons in <think> block before responding. Useful for complex scenes or multi-character interactions.
• No-think. Responds immediately with an empty <think> block. Better for fast, fluid exchanges.

Usage

vLLM

vllm serve Ks01/UwU-Qwen3.5-27B-v0.1 \
  --trust-remote-code \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.90

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Ks01/UwU-Qwen3.5-27B-v0.1",
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("Ks01/UwU-Qwen3.5-27B-v0.1")

messages = [
    {"role": "user", "content": "Write a tense reunion scene between two old friends at a rainy bus stop."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

• Experimental release (v0.1). Quality may be inconsistent.
• Context consistency can degrade in long multi-turn conversations.
• Thinking mode occasionally falls into analytical patterns instead of creative reasoning.

Credits

Fine-tuned from ArliAI/Qwen-3.5-27B-Derestricted, based on Qwen/Qwen3.5-27B.

License

Apache 2.0, inherited from Qwen3.5.