NaNovel: Chain-of-Thought Enhanced Creative Writer

Base: Qwen/Qwen3-VL-32B-Instruct
Context Length : Fine-Tuned with 128k context length.

Technical Summary

NaNovel is an experimental fine-tune designed to address a specific deficiency in current LLMs: narrative incoherence in long-context generation.

While most creative writing models focus on style transfer (prose quality), they often suffer from "hallucinated logic" writing beautiful sentences that contradict previous plot points. This model forces a Reasoning-First approach. By leveraging the Qwen3-VL architecture, we also introduce multimodal context awareness, allowing the model to ground its narrative logic in visual references if provided.

Trained with a 3:1 ratio of Reasoning to Creative data to instill structured "thinking" habits before narrative generation.

Methodology & Dataset

This model was fine-tuned using a carefully curated mix of datasets designed to foster long-context coherence, detailed reasoning, and authentic character voices. The breakdown is as follows:

Effective Mix

Global Distribution (Effective):

• Reasoning/Logic (~74%): The core of the model. Forces explicit planning.
• Creative/Context (~26%): Narrative flow, character voice, and worldbuilding.

Dataset Source	Focus	Weight (Norm)	Rationale
Reasoning Engaging Story	Logic-driven Storytelling	28%	Stories with explicit reasoning chains.
Qwill RP Creative Writing	Prose Quality & Roleplay	23%	Roleplay data enriched with creative reasoning.
Magpie Reasoning V2	Pure Chain-of-Thought	23%	Distilled reasoning from DeepSeek-R1.
LongPage	Extremely Long Context	7%	Chapter-length document handling.
Big5 Chat	Personality Consistency	6%	Dialogues based on Big Five traits.
Mental Health Counseling	Empathy & Psychology	5%	Realistic emotional depth.
LongForm	Long-form Generation	4%	General long-text capabilities.
World Building	Lore & Setting	2%	Detailed world descriptions.
BookSum + SimpleStories	Summarization & Flow	2%	Plot density and narrative coherency.

Qualitative Benchmarks & Samples

The following outputs were generated using a Temperature of 0.7 with a focus on logical consistency and sensory detail.

Check all examples on here

🔍 Reasoning & Plotting

Prompt: Write a mystery where the murder weapon was a block of ice that has since melted. The detective must use deductive reasoning to link the puddle of water to the killer. Outline the clues and the logical leap required.

Read the Model Output (Excerpt)

"Thorne approached it slowly, his breath catching. The sculpture was perfect, yet it radiated an unnatural chill. He knelt beside it, running his hand over its base. The marble was cold, but the ice itself felt unnervingly warm...

He looked up at the sculpture. It was clear, pure ice, yet it held within it a hidden warmth, a secret heat source. And now, the tiny droplets at its base were beginning to freeze, creating a miniature, crystalline barrier against the stone. The logical leap was terrifyingly clear: the ice sculpture wasn't just decorative; it was a weapon. The 'thud' Pettigrew heard was the impact of the sculpture, dropped from above...

High up, near the ceiling, a small, ornate vent was visible. He climbed onto a nearby bench, reaching up. His fingers brushed against the metal grate. It was warm to the touch... revealing a small, intricate mechanism inside – a tiny, hidden heating element..."

🌌 Atmosphere & Style

Prompt: Describe a beautiful but deadly alien jungle. Focus on the vibrant colors and the underlying sense of danger. Use lush, descriptive language.

Read the Model Output (Excerpt)

"The air in the Zephyr Jungles hung thick and heavy, a humid embrace that clung to the skin like a second layer. Sunlight, fractured by an impossibly dense canopy, dappled the forest floor in shifting patterns of emerald, sapphire, and gold...

But beneath this vibrant veneer lay a world of silent, predatory menace. The air, while sweet with the scent of exotic blossoms and decaying matter, carried an undercurrent of something sharper, more primal. It was the smell of life, yes, but life that was fiercely competitive...

A vine, thick as her wrist and adorned with jewel-like flowers of crimson and violet, might seem harmless, but its tendrils could coil around an unwary traveler with terrifying speed, constricting like a serpent. A patch of moss, soft and inviting underfoot, could conceal a nest of blind, venomous spiders, their tiny bodies camouflaged perfectly against the green."

🎭 Character Voice

Prompt: Write a monologue for a villain explaining their plan.

Read the Model Output (Excerpt)

"Ah, my dear, unsuspecting subjects of this once-great kingdom... I can almost taste the tension in the air, thick as the fog that rolls in from the sea every night. You think I am mad? Foolish? A mere dreamer with delusions of grandeur? Let me tell you, my friends, that I am neither. I am a visionary...

Your rulers, those who wear crowns of gold and jewels, have lulled you into a false sense of security... I have discovered a way to harness the power of the elements themselves. The wind, the water, the fire, and the earth...

So, will you stand with me? Will you join me in this quest for a brighter future? Or will you turn away, choosing instead to cling to the familiar, even as it leads you down a path of destruction?"

🚀 Sci-Fi Narrative

Prompt: Continue this story: 'The spaceship's AI spoke with a tremble in its voice. "Captain, I'm afraid I can't let you open that door. What's on the other side isn't... biological."...'

Read the Model Output (Excerpt)

"A chill ran down Captain Aris Thorne's spine as he stared at the reinforced door of the spaceship's medical bay. The AI... had never sounded so uncertain before... ‘What could possibly be worse than whatever caused us to lose contact with home?’

Just as he did so, however, a sudden burst of light erupted from within the medical bay... In the glare of the blinding illumination, figures began to take shape - tall, slender beings with elongated limbs and glowing eyes...

It was then that Aris realized just how wrong the AI had been - these creatures weren't merely non-biological; they were something else entirely. As the remaining crew members watched in horror, Aris found himself unable to break free from the alien's grasp. With each passing moment, he felt himself being drawn deeper into their world..."

Usage & Inference

This model is optimized for Multimodal & Text-Only workflows. It effectively utilizes the Qwen3-VL vision encoder.

1. Unsloth (Native High-Performance)

We recommend using FastVisionModel to leverage dynamic quantization if VRAM is constrained.

from unsloth import FastVisionModel
import torch

model, tokenizer = FastVisionModel.from_pretrained(
    model_name = "Dxniz/NaNovel",
    load_in_4bit = True,
)
FastVisionModel.for_inference(model)

system_prompt = "You are an expert novelist. always plan your narrative arc step-by-step before writing."
user_prompt = "Write a scene where a detective realizes the murder weapon was made of ice."

messages = [
    {"role": "system", "content": [{"type": "text", "text": system_prompt}]},
    {"role": "user", "content": [{"type": "text", "text": user_prompt}]}
]

input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
inputs = tokenizer(
    image=[], # Optional: Add image path here
    text=[input_text],
    add_special_tokens=False,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    **inputs, 
    max_new_tokens=8192, 
    use_cache=True, 
    temperature=0.7
)
print(tokenizer.batch_decode(outputs)[0])

2. HuggingFace Transformers

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "Dxniz/NaNovel",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("Dxniz/NaNovel")