Comparison of MinerU, Granite-Docling, and PaddleOCR-VL for PDF Parsing
A developer conducted a detailed benchmark comparing three open-source document parsing tools—MinerU, Granite-Docling, and PaddleOCR-VL—across 12 capabilities using six different document types on an L4 GPU. The test evaluated their performance on complex layouts like multi-column papers, scanned invoices, multi-lingual datasheets, and embedded charts. Document parsing is a critical bottleneck in Retrieval-Augmented Generation (RAG) pipelines, and this comparison provides practical, hands-on insights to help developers choose the right tool for specific document structures. Understanding how these models handle edge cases like footers, multi-lingual text, and tables is essential for building robust document AI workflows. Granite-Docling was the only tool that generated markdown-native pipe tables and preserved real heading levels, making its raw output the most readable. Meanwhile, MinerU successfully converted a bar chart into a table but silently discarded "page furniture" like footers (e.g., dropping an IBAN), requiring a custom rebuild from its block list to retain this data.
## BACKGROUND
Document parsing tools convert complex, unstructured PDFs and images into structured formats like Markdown or JSON, which are crucial for training AI models and feeding data into RAG systems. Granite-Docling is a compact vision-language model (VLM) developed by IBM, MinerU is an AI-native document parser designed for clean Markdown extraction, and PaddleOCR-VL is a resource-efficient VLM from Baidu's PaddlePaddle ecosystem.