---
license: agpl-3.0
language:
- zh
- en
pipeline_tag: image-text-to-text
library_name: transformers
---
MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing
[](https://github.com/opendatalab/MinerU/)
[](https://huggingface.co/opendatalab/MinerU2.5-2509-1.2B)
[](https://modelscope.cn/models/OpenDataLab/MinerU2.5-2509-1.2B)
[](https://huggingface.co/spaces/opendatalab/MinerU)
[](https://www.modelscope.cn/studios/OpenDataLab/MinerU)
---
# Introduction
**MinerU2.5** is a 1.2B-parameter vision-language model for document parsing that achieves state-of-the-art accuracy with high computational efficiency. It adopts a two-stage parsing strategy: first conducting efficient global layout analysis on downsampled images, then performing fine-grained content recognition on native-resolution crops for text, formulas, and tables. Supported by a large-scale, diverse data engine for pretraining and fine-tuning, MinerU2.5 consistently outperforms both general-purpose and domain-specific models across multiple benchmarks while maintaining low computational overhead.
## Key Improvements
- **Comprehensive and Granular Layout Analysis:** It not only preserves non-body elements like headers, footers, and page numbers to ensure full content integrity, but also employs a refined and standardized labeling schema. This enables a clearer, more structured representation of elements such as lists, references, and code blocks.
- **Breakthroughs in Formula Parsing:** Delivers high-quality parsing of complex, lengthy mathematical formulae and accurately recognizes mixed-language (Chinese-English) equations.
- **Enhanced Robustness in Table Parsing:** Effortlessly handles challenging cases, including rotated tables, borderless tables, and tables with partial borders.
# Quick Start
For convenience, we provide `mineru-vl-utils`, a Python package that simplifies the process of sending requests and handling responses from MinerU2.5 Vision-Language Model. Here we give some examples to use MinerU2.5. For more information and usages, please refer to [mineru-vl-utils](https://github.com/opendatalab/mineru-vl-utils/tree/main).
📌 We strongly recommend using vllm for inference, as the `vllm-async-engine` can achieve a concurrent inference speed of **2.12 fps** on one A100.
## Install packages
```bash
# For `transformers` backend
pip install "mineru-vl-utils[transformers]"
# For `vllm-engine` and `vllm-async-engine` backend
pip install "mineru-vl-utils[vllm]"
```
## `transformers` Example
```python
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from PIL import Image
from mineru_vl_utils import MinerUClient
# for transformers>=4.56.0
model = Qwen2VLForConditionalGeneration.from_pretrained(
"opendatalab/MinerU2.5-2509-1.2B",
dtype="auto", # use `torch_dtype` instead of `dtype` for transformers<4.56.0
device_map="auto"
)
processor = AutoProcessor.from_pretrained(
"opendatalab/MinerU2.5-2509-1.2B",
use_fast=True
)
client = MinerUClient(
backend="transformers",
model=model,
processor=processor
)
image = Image.open("/path/to/the/test/image.png")
extracted_blocks = client.two_step_extract(image)
print(extracted_blocks)
```
## `vllm-engine` Example (Recommended!)
```python
from vllm import LLM
from PIL import Image
from mineru_vl_utils import MinerUClient
from mineru_vl_utils import MinerULogitsProcessor # if vllm>=0.10.1
llm = LLM(
model="opendatalab/MinerU2.5-2509-1.2B",
logits_processors=[MinerULogitsProcessor] # if vllm>=0.10.1
)
client = MinerUClient(
backend="vllm-engine",
vllm_llm=llm
)
image = Image.open("/path/to/the/test/image.png")
extracted_blocks = client.two_step_extract(image)
print(extracted_blocks)
```
## `vllm-async-engine` Example (Recommended!)
```python
import io
import asyncio
import aiofiles
from vllm.v1.engine.async_llm import AsyncLLM
from vllm.engine.arg_utils import AsyncEngineArgs
from PIL import Image
from mineru_vl_utils import MinerUClient
from mineru_vl_utils import MinerULogitsProcessor # if vllm>=0.10.1
async_llm = AsyncLLM.from_engine_args(
AsyncEngineArgs(
model="opendatalab/MinerU2.5-2509-1.2B",
logits_processors=[MinerULogitsProcessor] # if vllm>=0.10.1
)
)
client = MinerUClient(
backend="vllm-async-engine",
vllm_async_llm=async_llm,
)
async def main():
image_path = "/path/to/the/test/image.png"
async with aiofiles.open(image_path, "rb") as f:
image_data = await f.read()
image = Image.open(io.BytesIO(image_data))
extracted_blocks = await client.aio_two_step_extract(image)
print(extracted_blocks)
asyncio.run(main())
async_llm.shutdown()
```
# Model Architecture
# Performance on OmniDocBench
## Across Different Elements
## Across Various Document Types
# Case Demonstration
## Full-Document Parsing across Various Doc-Types
## Table Recognition
## Formula Recognition
# Acknowledgements
We would like to thank [Qwen Team](https://github.com/QwenLM), [vLLM](https://github.com/vllm-project/vllm), [OmniDocBench](https://github.com/opendatalab/OmniDocBench), [UniMERNet](https://github.com/opendatalab/UniMERNet), [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR), [DocLayout-YOLO](https://github.com/opendatalab/DocLayout-YOLO) for providing valuable code and models. We also appreciate everyone's contribution to this open-source project!
# Citation
If you find our work useful in your research, please consider giving a star ⭐ and citation 📝 :
```BibTeX
@misc{niu2025mineru25decoupledvisionlanguagemodel,
title={MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing},
author={Junbo Niu and Zheng Liu and Zhuangcheng Gu and Bin Wang and Linke Ouyang and Zhiyuan Zhao and Tao Chu and Tianyao He and Fan Wu and Qintong Zhang and Zhenjiang Jin and others},
year={2025},
eprint={2509.22186},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.22186},
}
```