NaNovel-27B

NaNovel-27B is the main large autoregressive model in the Novelist series. It is designed as the balanced flagship for users who want stronger prose control, better narrative consistency, and more reliable instruction following than the 9B model without moving to a sparse Mixture-of-Experts architecture.

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-27B was fine-tuned on Dxniz/Novelist-CoT for creative writing, literary transformation, stylistic analysis, and reasoning-heavy language tasks. The training setup in this repository uses a long-context supervised fine-tuning pipeline with explicit planning behavior, allowing the model to reason about structure, emotion, pacing, and voice before writing the answer itself.

Within the lineup, this is the model to choose when output quality is the priority and standard dense-transformer inference is still preferred. It is especially good at high-control scene writing, consistent voice work, and prompts that mix literary output with explanation.

Evaluation Snapshot

Evaluation

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

The repository evaluation summaries show the following results for NaNovel-27B:

Overall evaluation results:

Detailed evaluation results:

This is the strongest evaluated autoregressive model in the current repository summaries. Its profile is notably balanced: it performs at a high level across prose, rewriting, translation, worldbuilding, emotional continuity, and craft-sensitive language tasks.

Recommended Use

Long-form fiction drafting
High-control scene writing with strict genre and style requirements
Rewriting prose while preserving voice
Prompted literary analysis and editorial explanation
Writers who want one general-purpose Novelist model with the best measured balance

Limitations

Larger hardware footprint than NaNovel-9B
Still requires prompt discipline for very long multi-chapter planning
Can over-elaborate if temperature and max token settings are too high
Generated text should be reviewed for factual claims, continuity, and rights-sensitive material

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Dxniz/NaNovel-27B"

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 tense literary scene in which two sisters meet at their mother's funeral."},
]

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

outputs = model.generate(
    inputs,
    max_new_tokens=1600,
    temperature=0.75,
    top_p=0.9,
    do_sample=True,
)

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