NaNovel-9B

NaNovel-9B is the smallest autoregressive release in the Novelist lineup. It is built for users who want the Novelist writing style, structured prewriting, and reliable creative task handling in a size that is easier to run than the larger series models.

Novelist Series

Base models: Qwen3.5-9B, Qwen3.5-27B, Qwen3.5-35B-A3B
Autoregressive models: NaNovel-9B, NaNovel-27B, NaNovel-35B-A3B
Diffusion models coming soon.

Model Overview

NaNovel-9B was fine-tuned on Dxniz/Novelist-CoT, a creative writing dataset centered on long-form prose, narrative planning, scene construction, stylistic control, and language-heavy editorial tasks. The model is intended to think through literary intent before producing final text, which makes it useful for drafting fiction, rewriting passages, analyzing tone, and answering craft-focused prompts.

Compared with the larger NaNovel variants, this model is the practical option for faster iteration, lighter hardware targets, and everyday writing assistance. It is best used when turnaround speed matters more than absolute depth.

Evaluation Snapshot

Evaluation

This model was evaluated with the Dxniz/Novelist-Bench benchmark dataset.

The repository evaluation summaries show the following results for NaNovel-9B:

Overall evaluation results:

Detailed evaluation results:

These numbers indicate that NaNovel-9B is strongest on narrative craft, style and voice, worldbuilding, and emotionally grounded prose. It is less dependable than the larger models on structure-heavy plotting, multilingual work, and translation-oriented tasks, which is consistent with its smaller scale.

Recommended Use

Short stories, scene drafts, and chapter starts
Style imitation with explicit voice constraints
Literary rewrites with explanation
Brainstorming character beats, imagery, and mood
Writing-adjacent language tasks such as translation commentary or craft analysis

Limitations

Long-range structural control is weaker than in NaNovel-27B
Plot logic and multi-part narrative architecture may need stronger prompting
Output quality can vary more on difficult multilingual or theory-heavy prompts
As with other instruction-tuned creative models, generated text should be reviewed before publication

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Dxniz/NaNovel-9B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are Novelist, a creative writing assistant."},
    {"role": "user", "content": "Write a gothic opening scene set in an abandoned observatory."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    return_tensors="pt",
    add_generation_prompt=True,
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=1200,
    temperature=0.8,
    top_p=0.9,
    do_sample=True,
)

print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))