🇨🇳 中文说明 | 🇺🇸 English

MiMo-VL-7B-RL Quick API Script Guide

The MiMoVLM-api_server.py script in this project is designed for rapid deployment and invocation of Xiaomi's open-source multi-modal Vision-Language Model (VLM) — MiMo-VL-7B-RL. It supports image captioning and other multi-modal reasoning tasks.

Model Overview

MiMo-VL-7B-RL is a high-performance vision-language model released by Xiaomi's large model team. It features powerful image understanding, reasoning, and multi-modal dialogue capabilities. The model utilizes a native-resolution ViT encoder, MLP projector, and MiMo-7B language model, and is optimized through multi-stage pre-training and mixed reinforcement learning, achieving state-of-the-art results on several public benchmarks.

• Model Homepage & Technical Report: Xiaomi MiMo-VL GitHub
• Model Weights Download: ModelScope Download Link

Main Features

• Supports image captioning via image URL or local file upload
• Customizable prompt (instruction) support
• Standard RESTful API interface for easy integration
• Automatic management of temporary files, suitable for high concurrency

Quick Start

1. Prepare Model Weights

   • Download the model weights from ModelScope Download Page and extract them to the directory specified by MODEL_PATH in MiMoVLM-api_server.py (default: /hy-tmp/data/MiMo-VL-7B-RL).

2. Install Dependencies

   pip install -r requirements.txt

3. Start the API Service

   python MiMoVLM-api_server.py

   The service will listen on http://0.0.0.0:8000 after startup.

API Endpoints

1. Image URL Captioning

• Endpoint: POST /describe_url/
• Request Body:

  {
    "image_url": "URL of the image",
    "prompt_text": "(Optional) Custom prompt"
  }

• Response Example:

  {
    "description": "Image caption",
    "prompt_used": "Prompt actually used",
    "error": null
  }

2. Image File Upload Captioning

• Endpoint: POST /describe_upload/
• Request Body:
  • image: The uploaded image file (form-data)
  • prompt_text: (Optional) Custom prompt
• Response: Same as above

Dependencies

See requirements.txt for details.

Acknowledgement

This project is based on the open-source MiMo-VL project by Xiaomi Large Model Team. Special thanks!

For more technical details, please refer to the MiMo-VL Technical Report.