Docling vs Marker: Battle of the Open-Source Document Parsers
Marker by VikParuchuri and Docling by IBM Research represent two different philosophies of PDF-to-Markdown conversion. Which one should you build with?
Marker is excellent for linear books and technical manuals, but IBM Docling dramatically outperforms on enterprise tables (TableFormer) and diverse multi-format ingestion (DOCX, PPTX, HTML).
1. Architectural Differences
Marker
Combines heuristics with deep learning models (Surya for layout and layout-ordering, Nougat-style formula detection). Highly tuned for arXiv papers and scientific books.
IBM Docling
Modular pipeline combining DocLayNet layout segmentation with TableFormer. Backed by corporate IBM enterprise research with rigorous validation on business and financial docs.
2. Table Extraction Capabilities
Tables are the ultimate test for any document AI. While Marker can parse simple bordered tables, it often breaks down on multi-tier financial balance sheets, cash flow statements, and cross-column footnotes.
Docling's TableFormer specifically predicts cell spans and topological relations, producing accurate TSV spreadsheets that can be copied directly into Excel via SheetParser.
3. Memory & GPU Footprint
Both tools require substantial GPU memory when running self-hosted:
- Marker: Requires ~6-8GB VRAM for smooth throughput. CPU mode is noticeably sluggish.
- Docling: Optimized for hybrid CPU/GPU execution. Fast on Apple Silicon (MPS) and NVIDIA T4/A10G.
If you don't have dedicated GPU hardware, you can run Docling via SheetParser with zero installation.