Open-Source PDF Parser Benchmark: MinerU, Granite‑Docling, PaddleOCR‑VL, and XBerg

The author revisits a Reddit post that performed a side‑by‑side comparison of four open‑source PDF parsers on a unified sample set of six document types, running the three VLM‑based tools on an L4 GPU and XBerg on a CPU.

Evaluation Method and Test Documents

Six representative PDFs were selected: a multi‑row header financial report, a two‑column arXiv paper, a German scanned invoice without text layer, a French municipal report with an embedded bar chart, a multilingual specification page, and a three‑column newsletter.

Tool Overview and Speed

MinerU 2.5 – 1.2 B parameters, L4 GPU, 4.7 s per page.

Granite‑Docling – 258 M parameters, L4 GPU, 2.8 s per page.

PaddleOCR‑VL – 0.9 B parameters, L4 GPU, 3.6 s per page.

XBerg 1.0 – CPU only, 0.2 s per page.

Capability Matrix (12 dimensions)

Simple tables: MinerU ✓, Docling ✓, Paddle ✗, XBerg ✓

Merged headers: MinerU ✓, Docling ≈, Paddle ✗, XBerg ✗

Mathematical formulas: MinerU ✓, Docling ≈, Paddle ✗, XBerg ✗

Scanned pages: MinerU ✓, Docling ≈, Paddle ✓, XBerg ≈

Small fonts / footers: MinerU ✓, Docling ✗, Paddle ✓, XBerg ✓

Title hierarchy: MinerU ≈, Docling ≈, Paddle ✗, XBerg ≈

Column order: MinerU ✓, Docling ✓, Paddle ≈, XBerg ✗

Image captions: MinerU ✓, Docling ✓, Paddle ✓, XBerg ≈

Chart data: MinerU ✓, Docling ✗, Paddle ✗, XBerg ✗

Multilingual text: MinerU ✓, Docling ✓, Paddle ✓, XBerg ≈

Numeric formatting: MinerU ✓, Docling ✓, Paddle ✓, XBerg ✓

Pagination handling: MinerU ✓, Docling ✓, Paddle ✓, XBerg ✓

Tool‑Specific Findings

MinerU passes 11 of 12 items; the only downgrade is title handling (hash symbols are promoted). It excels at tables, formulas, scans, and extracting values from bar charts, but costs 4.7 s per page and may drop footers in default markdown output.

Granite‑Docling is the fastest VLM (2.8 s per page) and produces the cleanest markdown for well‑structured documents, yet it struggles with scanned pages, loses footers, and retains only figure titles for charts.

PaddleOCR‑VL shows strong OCR accuracy on scans and footers, but its structural parsing fails on simple tables, merged headers, formulas, and titles, behaving more like a pure OCR engine.

XBerg offers extreme throughput (0.2 s per page) on CPU, but its ability to handle merged headers, formulas, multi‑column layouts, and charts is weak; it is suitable for quick pre‑filtering of simple documents.

Common Weaknesses

All tools have difficulty extracting numeric values from chart figures; only MinerU reads bar‑chart values, while the others stop at the figure title.

Title hierarchy is often mis‑detected, with citations, authors, and dates being promoted into headings, which can pollute downstream RAG chunking.

Practical Recommendations

Prioritize accuracy: MinerU 2.5 for the most complete extraction.

Prefer clean markdown output: Granite‑Docling for well‑formatted PDFs.

Use PaddleOCR‑VL when OCR robustness is needed and you can post‑process structure.

Choose XBerg for CPU‑only, high‑speed screening of simple pages.

For chart value extraction, start with MinerU and verify with a dedicated model.

Other Considerations

Real‑world documents may contain handwriting, blurry scans, stains, broken sections, unusual tables, emojis, or icons, which can cause failures even if the benchmark looks good.

There is no single “silver‑bullet” parser; the optimal solution depends on document type and operational constraints such as speed, quality, and cost.