Docling vs MinerU.
Both modern PDF extraction tools built for complex documents. MinerU holds the top spot on layout detection benchmarks with 97.5 mAP. Docling focuses on speed and simplicity. The question is whether benchmark performance translates to real-world usability.
The numbers, row by row.
| Metric | Docling | MinerU 2.5 |
|---|---|---|
| Time (12 pages) | 8.2s | 14.7s |
| Layout detection mAP | 93.1% | 97.5% |
| Table structure preserved | Partial | Full |
| Equation handling | Basic | LaTeX output |
| Figure extraction | Yes | Yes |
| Setup complexity | Simple | Medium |
| License | Apache 2.0 | AGPL-3.0 |
Test PDF. 12 pages, academic paper with tables, figures, and equations.
Docling output — fast, structure partially lost
Detected all major elements. Tables were detected but the internal structure was partially lost.
[TITLE] Deep Learning for Document Understanding
[AUTHOR] Smith et al.
[ABSTRACT] We propose a novel approach...
[SECTION] 1. Introduction
[TEXT] Document understanding has become...
[TABLE]
Detected but structure partially lost
[FIGURE_CAPTION] Figure 1: Architecture overview
...MinerU 2.5 output — structure preserved
Tables preserved with proper markdown formatting. Equations converted to LaTeX. The structure is production-ready.
[title] Deep Learning for Document Understanding
[author] Smith et al.
[abstract] We propose a novel approach...
[section] 1. Introduction
[text] Document understanding has become...
[table]
| Model | Accuracy | Speed |
|-------|----------|-------|
| Ours | 94.2% | 12ms |
| SOTA | 92.1% | 18ms |
[figure_caption] Figure 1: Architecture overview
...Get the full OCR comparison spreadsheet
30+ models × 8 benchmarks, accuracy + price per page. We email it and keep it current.
When to pick which.
- Table accuracy is critical
- Academic or technical PDFs
- You need LaTeX equations
- Structure preservation matters more than speed
- You can live with AGPL-3.0
- Speed matters (~2x faster)
- Simple documents
- You need an Apache 2.0 license
- Basic extraction is sufficient
- You’re distributing commercial software
For research papers, financial reports, or any document where table and formula accuracy matters, MinerU’s extra processing time is worth it. Docling uses Apache 2.0, which is permissive for commercial use; MinerU uses AGPL-3.0, which requires source release if you distribute. For SaaS (no distribution), both are fine.
The code.
Docling
from docling import Docling
docling = Docling()
result = docling.parse('research_paper.pdf')
for page in result.pages:
for element in page.elements:
print(f"{element.type}: {element.text}")MinerU
from mineru import MinerU
miner = MinerU()
result = miner.extract('research_paper.pdf')
for page in result:
for block in page.blocks:
print(f"[{block.category}] {block.text}")
if block.category == 'table':
print(block.to_markdown())Adjacent comparisons.
Get the full OCR comparison spreadsheet
30+ models × 8 benchmarks, accuracy + price per page. We email it and keep it current.