Qwen3.5-Hasutsubomi-9B (Pre-Alpha)

Qwen3.5-Hasutsubomi-9B is a vision-language fine-tuned large language model based on the Qwen3.5-9B architecture. It is specifically tailored to act as an expert in Japanese creative writing and little sub-culture knowledge (including anime and manga).

Note: This model is currently in a pre-alpha state. It is still undergoing training and experimentation.

About the Name

Hasutsubomi (蓮蕾) translates to "Lotus Bud" in Japanese. This name reflects the model's current pre-alpha status—a beautiful flower that is still growing and gathering knowledge. Once the model completes its training cycle and reaches maturity, it is planned to evolve into Hanahiraki (Blooming).

Model Highlights

• Japanese Sub-Culture Expert: Little knowledge of Japanese internet lore, anime tropes, and light novel storytelling structures.
• Creative Writing Focus: Fine-tuned to generate better-quality prose, character dialogues, and world-building outlines.
• Keep Vision-Language Capabilities: Inherited from the Qwen3.5 base, Hasutsubomi can (hopefully) analyze anime illustrations, manga panels and write stories or descriptions based on them.
• Pre-Alpha State: This is an early checkpoint release. Expect highly creative but occasionally unstable outputs as the model continues to learn.
• Efficient Architecture: Powered by Qwen3.5's Gated Delta Networks and sparse Mixture-of-Experts for high-throughput generation.

Model Overview

• Type: Causal Language Model with Vision Encoder
• Base Model: Qwen/Qwen3.5-9B
• Fine-tuning Stage: Pre-Alpha (Work in Progress)
• Primary Languages: Japanese, English
• Number of Parameters: 9B
• Hidden Dimension: 4096
• Number of Layers: 32
• Context Length: Natively supports up to 262,144 tokens (excellent for long-form light novel generation or context-heavy RP).

Intended Use

Direct Use Cases

• Story Generation: Writing light novel chapters, fanfiction, or visual novel scripts.
• Image-to-Story: Uploading a character illustration and asking the model to write a backstory or describe the character's design.
• Cultural Explanation: Explaining complex or niche Japanese anime history, or manga tropes. (limited performance for now)

Limitations & Biases (Pre-Alpha)

• As a pre-alpha checkpoint, the model may occasionally suffer from formatting instability or repetitive phrasing in extended writing tasks.
• It might heavily lean into dramatic or stereotypical anime tropes even when a neutral tone is requested.
• Hallucinations may occur when asked about highly obscure, real-world sub-culture events.

Quickstart

Because Qwen3.5 models operate in a "thinking mode" by default, Hasutsubomi will often draft its creative thoughts or plot outlines before writing the final response.

Below is a guide to using the model via the OpenAI-compatible API format (assuming you are serving the model via vLLM, SGLang, or standard Transformers).

Creative Writing (Text-Only)

from openai import OpenAI
# Configured by environment variables
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

messages =[
    {"role": "system", "content": "You are a professional Japanese light novel author."},
    {"role": "user", "content": "Write a short prologue for a fantasy Isekai story where the protagonist is reincarnated not as a hero, but as a magical library. Use an atmospheric and slightly melancholic tone in Japanese."}
]

chat_response = client.chat.completions.create(
    model="Local-Novel-LLM-project/Qwen3.5-Hasutsubomi-9B",
    messages=messages,
    max_tokens=4096,
    temperature=0.8, # Slightly higher temperature for creative writing
    top_p=0.9,
    presence_penalty=1.5,
)
print("Chat response:", chat_response.choices[0].message.content)

Analyzing Anime/Manga Art (Vision-Language)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

messages = [
    {
        "role": "user",
        "content":[
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://example.com/anime_character_design.jpg"
                }
            },
            {
                "type": "text",
                "text": "Analyze this character's design. Based on typical Japanese sub-culture tropes, what kind of personality (e.g., Tsundere, Yandere, Chuunibyou) do they likely have, and what kind of magic would they use in a fantasy setting?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Local-Novel-LLM-project/Qwen3.5-Hasutsubomi-9B",
    messages=messages,
    max_tokens=2048,
    temperature=0.7,
)
print("Chat response:", chat_response.choices[0].message.content)

Best Practices for Hasutsubomi

1. Prompting for Writing: When asking the model to write stories, it helps to explicitly state the desired trope or genre (e.g., "Cyberpunk", "Isekai", "Slice of Life", "Iyashikei").
2. Sampling Parameters for Creativity:
   • For creative writing and roleplay, we recommend a higher temperature: temperature=0.8 to 1.0 with a top_p=0.9.
   • Apply a slight presence_penalty=1.5 to encourage varied vocabulary.
3. Thinking Mode: Allow the model to utilize <think>\n...\</think> tags to plan out story beats before it generates the final text. This significantly improves narrative coherence.

Citation

If you use this fine-tuned model, please consider acknowledging this repository. Additionally, please cite the original Qwen3.5 paper:

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}