Scored 10 November 2026
MarkItDown vs Docling for PDFs: what 200 test files show
MarkItDown vs Docling comes down to structure versus speed. On opendataloader-bench's 200 PDFs, Docling scored 0.882 overall and MarkItDown 0.589: MarkItDown wrote almost no headings and scored 0.273 on tables to Docling's 0.887, but its plain text was not far behind (0.844 against 0.898), and it took 0.11 seconds a PDF to Docling's 0.76.
A word on who wrote this: markdownwise is our own PDF to Markdown converter, which makes us a competitor of both. Neither program was run by us. The benchmark keeps the Markdown each one produced, and we put that through the benchmark's own checker, the same one that gives our numbers.
The numbers
| Score (0 to 1) | MarkItDown 0.1.5 | Docling 2.84.0 | markdownwise (ours) |
|---|---|---|---|
| Overall mean of the scores a PDF has |
0.589 | 0.882 | 0.873 |
| All text, in order 200 PDFs |
0.844 | 0.898 | 0.906 |
| Text that isn't in a table 200 PDFs |
0.838 | 0.855 | 0.900 |
| Tables 42 PDFs, cell words and grid |
0.273 | 0.887 | 0.825 |
| Table grid 42 PDFs, rows and columns only |
0.328 | 0.901 | 0.844 |
| Headings 107 PDFs, titles and levels |
0.000 | 0.824 | 0.793 |
| Plain-text PDFs 70 PDFs with no table or heading |
0.869 | 0.909 | 0.914 |
| Seconds per PDF benchmark timing, Apple M4 |
0.11 | 0.76 | see note |
A score of 1 means the Markdown matches the benchmark's hand-made answer key. The plain-text row takes the 70 PDFs that have neither a table nor a heading, where only the words and their order count. Our 0.94 seconds a PDF was timed on a Windows laptop, not the benchmark's machine, so it stays out of the speed row.
What MarkItDown leaves out
- Headings. Only 1 of MarkItDown's 200 files has a single # heading. Titles come out as ordinary lines, so the outline of a report or a paper is gone, and the headings score is 0 on all 107 PDFs that have headings.
- Table layouts. On 18 of the 42 PDFs with a table, its table score was 0. On a summary of land-ownership rules by country, its output broke the table into five separate pieces, with parts of the column titles left outside them; Docling gave one table with a row per country.
- The cloud option wasn't tested. This test used MarkItDown's built-in, offline PDF reader. Its README also describes an optional Azure Document Intelligence step for cloud layout extraction, which this test didn't use.
Where MarkItDown holds its own
- Words outside tables: a dead heat. MarkItDown was ahead on 60 PDFs, Docling on 60, and 80 were within 0.02.
- Contents pages: Docling turned two contents lists into tables, while the answer key keeps them as plain lines. MarkItDown kept the lines, though on one of the two it listed all the page numbers after the chapter titles.
- Speed and reach: about seven times quicker than Docling in the benchmark's runs, and the same call also reads Word, PowerPoint, Excel, HTML, EPUB and more. Its README says it is built for feeding text to language models and analysis pipelines rather than for faithful documents that people read.
Head to head, one PDF at a time
| Score | Docling ahead | MarkItDown ahead | Within 0.02 |
|---|---|---|---|
| Overall 200 PDFs | 140 | 21 | 39 |
| All text, in order 200 PDFs | 74 | 54 | 72 |
| Text that isn't in a table 200 PDFs | 60 | 60 | 80 |
| Tables 42 PDFs | 40 | 0 | 2 |
| Headings 107 PDFs | 105 | 0 | 2 |
Choosing between them
Text for a language model, from a mix of file types: MarkItDown. It reads PDFs and Office files through one interface and is quick; on plain-text PDFs it scored 0.869 here. Expect tables and headings to come through flattened.
Reports, papers or anything with tables: Docling. It led on 40 of the 42 table PDFs and 105 of the 107 with headings.
Licenses: both are MIT. MarkItDown comes from Microsoft; Docling began at IBM Research Zurich and is now hosted by the LF AI & Data Foundation. Both are Python packages you install.
No install at all: the free online PDF to Markdown converter here scored 0.873 on the same files, with its own engine rather than either of these: below Docling on tables and headings, ahead of both on text.
Where the numbers come from
opendataloader-bench is an open test of 200 document pages from Upstage's DP-Bench, each with a hand-made Markdown answer. It ran MarkItDown 0.1.5 and Docling 2.84.0 in April 2026 and published both outputs. We scored them with the benchmark's checker as of 2 November 2026; both match the benchmark's own stored results exactly. A program's newer release may do better or worse.
Also on this site: Marker and Docling on the same 200 PDFs, seven online converters put to one test, and what happens inside a PDF to Markdown conversion.
Questions about MarkItDown and Docling
Is MarkItDown or Docling better for PDFs?
For PDFs with tables or headings, Docling: it scored 0.882 overall to MarkItDown's 0.589 on the same 200 files. On PDFs that are only running text the gap is small, 0.909 to 0.869, and MarkItDown is quicker.
Why does MarkItDown get 0 for headings?
Its built-in PDF reader writes titles as ordinary lines: only 1 of its 200 files in this test had a # heading. The headings score compares the document's outline, so a file with no headings scores 0.
Is MarkItDown faster than Docling?
Yes, in the benchmark's runs on an Apple M4: 0.11 seconds per PDF for MarkItDown and 0.76 for Docling.
Can MarkItDown read tables in a PDF?
Sometimes. Across the 42 PDFs with a table its table score averaged 0.273, and it was 0 on 18 of them. Its README offers an optional Azure Document Intelligence step for PDFs, which this test didn't use.