mai_replier-9b-v3

mai_replier-9b-v3

Plays a persona in Chinese QQ group chats by reading the chat history and writing one short, casual reply with no visible reasoning. A merged 9B fine-tune of Qwen3.5-9B that also reads inline images.

Made by tcmofashi on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: mai_replier-9b-v3 is the replier model of the MaiBot ecosystem, a bot that joins QQ group chats in a fixed character. It reads the group chat history and writes one short, casual reply in character, with no chain of thought. It is a full merged 9B model fine-tuned from Qwen3.5-9B with LoRA, and it can read images placed inline in the chat.

Who it suits: People who run MaiBot-style chat bots in Chinese group chats and want a self-hosted model that replies in persona.

Industry | Who uses it | How they use it | Example
Community | Group admins | Run a persona bot in a QQ group | A character that chats with members
Software | Bot developers | Self-host the replier behind MaiBot | Serve it with vLLM for the bot
Research | Dialogue researchers | Study persona-conditioned group chat replies | Test how it handles images in the chat

If you want a general assistant or an English chat model, you can skip it.

Where to use it:

• QQ group chats, through the MaiBot ecosystem
• Your own computer with Python and transformers, on Windows, macOS or Linux
• A vLLM server, for hosting it behind a bot

Use case | Who | What they get back
Persona replies | Bot developers | One short, casual line in character
Chats with images | Group admins | Replies that take inline images into account
Self-hosted serving | Platform teams | An endpoint for the bot to call

How to set it up:

• Python with pip
• About 18.8 GB of disk space for the BF16 weights, or about 4.7 GB at 4-bit (an estimate)
• A GPU with enough memory, since the card's snippet uses device_map auto

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Load them with transformers, or start a vLLM server.

pip install -U huggingface_hub
hf download tcmofashi/mai_replier-9b-v3

How to use it: Build two messages. The system message holds the character's name and persona. The user message holds the chat history lines and a closing prompt asking the character to reply. Always pass enable_thinking as False. The sampling settings sit in the generate call below.

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

MODEL = "tcmofashi/mai_replier_9b_v3"

processor = AutoProcessor.from_pretrained(MODEL)
model = AutoModelForImageTextToText.from_pretrained(MODEL, dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "system", "content": "你的名字是小千。是一个大二在读女大学生…"},
    {"role": "user", "content": "你正在qq群里聊天…\n现在，你说："},
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True,
    chat_template_kwargs={"enable_thinking": False},
)
inputs = processor(text=text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, temperature=0.7, top_p=0.95,
                     repetition_penalty=1.15, do_sample=True, eos_token_id=[248046, 248044])
print(processor.tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

To serve it with vLLM, the card gives this command.

vllm serve tcmofashi/mai_replier_9b_v3 \
    --served-model-name mai_replier_9b_v3 \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.90 \
    --port 8200

Example: You give the model a persona and a short chat history where someone asks the character a question. It returns one casual sentence in that character's voice. This is an illustration of the input and output the card describes.

For newcomers and experts: If you are new, know that this model is built for one job, replying inside a chat in character. Start by copying the card's snippet and changing the persona text.

For experts, the card lists Qwen3.5-9B with 24 linear-attention GDN layers and 8 full-attention layers, LoRA SFT with r=64 and alpha=32, merged into full weights. Images are embedded where they appear in the text. The latest 3 keep a 640 long edge and older ones shrink to 128.

Chat history becomes a system and user message
Images are embedded at their place in the text
The model skips reasoning and replies
One short in-character line comes out

Good to know: The card calls this a community experiment for QQ group role play.

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need a GPU for a 9B model.
• Data it reaches: downloading reaches Hugging Face. The model only reads the chat text and images you pass in.
• Not for: speaking for real people. The card says outputs do not represent any real person's views, and its prompt tells the model to avoid illegal, sexual, violent and political content.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 9.4B
• Weights: about 18.8 GB in BF16; about 4.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 660
• Likes: 8
• License: apache-2.0
• Languages: zh
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 3
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, text-generation, chatbot, qq-group-chat, roleplay, chinese, retro-replyer, multimodal, interleaved-image

Links

• Model: https://huggingface.co/tcmofashi/mai_replier-9b-v3
• Files: https://huggingface.co/tcmofashi/mai_replier-9b-v3/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download tcmofashi/mai_replier-9b-v3

Then follow the usage in its model card. Nothing in this kit runs by itself.