opendatalab / MinerU2.5-2509-1.2B

MinerU2.5

We are releasing MinerU2.5, a 1.2B-parameter visual-language model specialized in OCR and document parsing, enabling more accurate and robust parsing of complex and diverse real-world documents.

The model weights are stable and available for use, primarily intended for internal development and demonstration purposes.

⚠️ A full technical report, source code, and a comprehensive README will be released later this month. Stay tuned!

Quick Start

For convenience, we provide a python package named mineru-vl-utils to smoothly use the MinerU2.5 Vision-Language Model. For more information and usages, please refer to mineru-vl-utils .

Here we give a simple example to use MinerU2.5 with 🤗 Transformers.

Install packages

pip install mineru-vl-utils[transformers]

Run with Transformers

from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from PIL import Image
from mineru_vl_utils import MinerUClient

model_path = "opendatalab/MinerU2.5-2509-1.2B"

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_path, 
    dtype="auto",
    device_map="auto"
)

processor = AutoProcessor.from_pretrained(
    model_path,
    use_fast=True
)

client = MinerUClient(
    backend="transformers",
    model=model,
    processor=processor
)

image_path = '/path/to/your/image'
image = Image.open(image_path)
extracted_blocks = client.two_step_extract(image)