Scored 9 November 2026
Marker vs Docling: the same 200 PDFs, scored side by side
Marker vs Docling, in numbers: on the 200 PDFs of opendataloader-bench, Docling scored 0.882 overall and Marker 0.867, where 1.0 means the Markdown matches the hand-made answer key. Docling read tables much better (0.887 against 0.808); Marker kept the words outside tables a little better. In the benchmark's own timing, Docling needed 0.76 seconds a PDF and Marker 53.9.
We make markdownwise, a converter of our own, so read this knowing we have a stake. We didn't run either program ourselves: the benchmark publishes the Markdown each one produced, and we scored it with the benchmark's own open-source checker. Our column is scored by that same checker.
Scores, measure by measure
| What is measured | Marker 1.10.1 | Docling 2.84.0 | markdownwise (ours) |
|---|---|---|---|
| Overall the average of the measures below that a PDF has |
0.867 | 0.882 | 0.873 |
| Text in reading order all 200 PDFs |
0.890 | 0.898 | 0.906 |
| Text outside tables all 200 PDFs, tables left out |
0.863 | 0.855 | 0.900 |
| Tables, words and layout the 42 PDFs with a table |
0.808 | 0.887 | 0.825 |
| Tables, layout only rows and columns, words ignored |
0.834 | 0.901 | 0.844 |
| Headings, words and levels the 107 PDFs with headings |
0.819 | 0.824 | 0.793 |
| Headings, levels only the outline's shape, words ignored |
0.936 | 0.906 | 0.869 |
| Seconds per PDF the benchmark's runs on an Apple M4 |
53.9 | 0.76 | see below |
Each score runs from 0 to 1. Text compares every word with the answer key in reading order; tables and headings count only on the PDFs that have them. markdownwise took 0.94 seconds a PDF on a Windows laptop's processor, a different machine, so that time isn't comparable with the other two.
PDF by PDF: how often each one came out ahead
An average can hide a lot, so we also compared Marker and Docling one PDF at a time. A lead of more than 0.02 counts as a win; anything closer counts as a tie.
| Compared on | Docling ahead | Marker ahead | Within 0.02 |
|---|---|---|---|
| Overall, all 200 PDFs | 62 | 37 | 101 |
| Text in reading order, 200 PDFs | 47 | 35 | 118 |
| Tables, 42 PDFs | 11 | 7 | 24 |
| Headings, 107 PDFs | 60 | 26 | 21 |
Docling's lead is real but not a sweep. Marker came out ahead on 37 of the 200 PDFs, and on 101 the two scored within 0.02 of each other.
Where each one slipped
- Docling lost the text of a poster-style page. On a one-page copyright fact sheet laid out like a poster, with pictures between the lines of text, Docling returned only picture placeholders. Marker kept the text.
- Docling ran a table into the paragraphs. On a fluid-mechanics page whose table of figures is printed as a picture below the text, Docling mixed the table's numbers into the sentences above it. Marker kept the paragraphs whole and the table as a table.
- Marker split wrapped table cells into rows. On a summary table with long notes in each cell, Marker started a new row for every wrapped line: 36 rows for 5 countries. Docling kept one row per country.
Which one fits your job
Lots of tables, or a big batch: Docling. It led on tables by a wide margin and was far quicker in the benchmark's runs. Its code is under the MIT license, and the project began at IBM Research Zurich and now sits with the LF AI & Data Foundation.
Equations you want as LaTeX: Marker's documentation says it writes them as LaTeX between
$$ marks, and it did so on the fluid-mechanics page here.
Commercial use: read the terms first. Marker's code is Apache-2.0, but its model weights are free only for research, personal use and startups under $5 million in funding or revenue; above that, its makers sell a license. Docling's README points to each model's own license for the models it uses.
One PDF, no install: both are Python programs you install and run yourself. If you only need a file converted, the free PDF to Markdown converter on this site scored 0.873 on the same 200 PDFs. It runs its own engine, not Marker or Docling, and was ahead of both on words outside tables but behind both on headings.
How the scores were made
The PDFs and answer key come from opendataloader-bench, an open benchmark of 200 pages from real documents, taken from Upstage's DP-Bench. Its checker scores text by edit distance, tables by tree edit distance (TEDS) and headings by the outline they form. The benchmark publishes each tool's Markdown, so anyone can score it again.
We scored Marker 1.10.1 (run in January 2026) and Docling 2.84.0 (run in April 2026) with the checker as it stood on 2 November 2026. Marker's result file in the benchmark lists 0.861, scored with an earlier version of the checker; the same Markdown scores 0.867 today, almost all of the change in headings. Newer releases of either program may score differently.
For a test of free websites rather than programs, see our test of seven online PDF to Markdown converters, and for what happens inside a conversion, how markdownwise reads a PDF. Microsoft's lighter tool is scored the same way in MarkItDown vs Docling.
Questions about Marker and Docling
Is Docling better than Marker?
On this benchmark, overall, yes: 0.882 to 0.867, mostly thanks to tables. But Marker came out ahead on 37 of the 200 PDFs and kept the words outside tables slightly better, so try both on a few of your own files.
Which is faster, Marker or Docling?
In the benchmark's published runs on an Apple M4, Docling averaged 0.76 seconds per PDF and Marker 53.9. Marker's README says it runs on a GPU, a CPU or Apple's MPS, and picks a faster mode by default on the last two.
Can I use Marker or Docling in a commercial product?
Docling's code is MIT-licensed. Marker's code is Apache-2.0, but its model weights are free only for research, personal use and startups under $5 million in funding or revenue; bigger companies need a paid license. Check each project's current terms before you ship.
Can I check these scores myself?
Yes. opendataloader-bench on GitHub publishes the 200 PDFs, the answer key, the Markdown each tool produced and the checker. Run the checker on the Marker and Docling folders and you get the numbers on this page.