MarkItDown vs Docling vs Marker vs Pandoc: an honest comparison

7 August 2026 · Milldown team · All posts

Four open-source projects dominate document-to-Markdown conversion, and they are genuinely different tools built for different jobs. Full disclosure up front: Milldown is built on MarkItDown, so we have a horse in this race — which is exactly why this comparison names the cases where the others win.

The one-table answer

MarkItDown: the generalist

Microsoft’s MarkItDown covers the widest range of everyday office and web formats of the four — Word, Excel, PowerPoint, PDF, HTML, EPUB, images and more — into Markdown specifically shaped for LLM consumption. (Pandoc reads more formats in total, around fifty, but not the one that matters most here: PDF.) It’s a small Python library with no model downloads and no GPU requirement, which makes it fast and easy to embed. The trade-off: it reads a PDF’s text layer rather than analysing its visual layout, so a gnarly multi-column report with nested tables will come out less faithfully than from Docling or Marker.

Docling: when layout is the hard part

IBM’s Docling runs vision transformer models over each page to detect columns, tables, figures, and reading order. Layout analysis is its design goal, and on scientific papers or financial reports with dense tables it is the one to reach for when structure is the hard part. The cost is speed and footprint: model inference per page versus text extraction. If your corpus is layout-heavy and accuracy is worth minutes, Docling earns its keep.

Marker: the academic specialist

Marker focuses on doing one thing extremely well: research PDFs. Equations, citations, and figure handling are its home turf, and it is designed to use a GPU where one is available. If you’re converting arXiv papers all day, start here. For mixed office documents, it’s the wrong shape.

Pandoc: the publisher (and the trap)

Pandoc is the finest universal document converter ever written, and its DOCX-to-Markdown output is arguably the best there is. But the number-one misconception in this space: Pandoc cannot read PDFs. PDF is an output format for Pandoc (via LaTeX), not an input. It also has no notion of token efficiency, OCR, or LLM-oriented cleanup — it’s a publishing tool from a different era, and a brilliant one. We wrote a fuller comparison in Milldown vs pandoc.

Which should you pick?

• Mixed documents into an AI workflow — MarkItDown. Breadth and speed win when the input is “whatever lands in the folder.”
• Dense multi-column PDFs where structure is everything — Docling.
• Research papers at volume — Marker.
• DOCX in, EPUB/LaTeX/anything out — Pandoc, no contest.

These are libraries, though — all four assume a terminal and a Python or Haskell toolchain. If you want the MarkItDown engine with none of the setup, that’s the product we build: Milldown wraps the full engine in a native Mac app and adds what the library doesn’t have — on-device OCR for scanned PDFs, a live token counter, boilerplate cleanup, and batch watch folders. Everything runs locally.