Result: Altto’s PDF to Word converter scored 0.884 out of 1 on 200 public PDFs (11 October 2026). That puts it level with Docling (0.882) and Nutrient (0.885), and behind OpenDataLoader’s hybrid mode (0.907). The others are developer tools; Altto is a free website with nothing to install, and your PDF never leaves your device.
The results
Each converter is scored on three things, each out of 1, where 1 is a perfect match with the hand-checked answer. Reading order checks that the text comes out complete and in the right order. Tables checks that tables come out as real tables with the right cells. Headings checks that headings are marked as headings, at the right level. Overall is the average.
| # | Converter | Overall | Reading order | Tables | Headings |
|---|---|---|---|---|---|
| 1 | OpenDataLoader (hybrid mode) Open source, Apache-2.0 | 0.907 | 0.934 | 0.928 | 0.821 |
| 2 | Nutrient Commercial SDK | 0.885 | 0.925 | 0.708 | 0.819 |
| 3 | Altto PDF to Word Free website (this site) | 0.884 | 0.906 | 0.862 | 0.831 |
| 4 | Docling (IBM) Open source, MIT | 0.882 | 0.898 | 0.887 | 0.824 |
| 5 | OpenDataLoader (hybrid, Hydrogen) Open source, Apache-2.0 | 0.877 | 0.926 | 0.796 | 0.769 |
| 6 | Marker Open source, GPL-3.0 | 0.861 | 0.890 | 0.808 | 0.796 |
| 7 | OpenDataLoader (hybrid, Helium) Open source, Apache-2.0 | 0.845 | 0.879 | 0.807 | 0.755 |
| 8 | Unstructured (hi-res) Open source, Apache-2.0 | 0.841 | 0.904 | 0.588 | 0.749 |
| 9 | EdgeParse Open source, Apache-2.0 | 0.837 | 0.894 | 0.717 | 0.706 |
| 10 | OpenDataLoader Open source, Apache-2.0 | 0.831 | 0.902 | 0.489 | 0.739 |
| 11 | MinerU Open source, AGPL-3.0 | 0.831 | 0.857 | 0.873 | 0.743 |
| 12 | PyMuPDF4LLM Open source, AGPL-3.0 | 0.732 | 0.885 | 0.401 | 0.412 |
| 13 | Unstructured (fast) Open source, Apache-2.0 | 0.686 | 0.882 | 0.000 | 0.388 |
| 14 | MarkItDown (Microsoft) Open source, MIT | 0.589 | 0.844 | 0.273 | 0.000 |
| 15 | LiteParse Open source, Apache-2.0 | 0.576 | 0.866 | 0.000 | 0.000 |
Reading order is scored on all 200 PDFs, tables on the 42 that contain tables, and headings on the 107 that contain headings. Scores for the other converters come from the opendataloader-bench repository (published April 2026); we did not re-run them. Download our per-document scores: pdf-to-word-benchmark.json.
What the numbers mean in practice
- 112 of the 200 PDFs scored 0.9 or more with Altto: text in order, tables intact, headings recognised. These are mostly reports, papers, letters and forms.
- 10 scored under 0.7. These are the pages to watch: posters set as one big picture, slides with text scattered across the page, and tables that are only a picture inside the PDF.
- Averages hide differences. Altto and Docling have almost the same average, but document by document Docling did better on 109 PDFs and Altto on 63. Altto’s strength is headings and complete text; Docling is more consistent across odd layouts.
How Altto has improved
We use this test to find and fix problems in the converter, and we re-run it after each round of changes. The scores below are for the whole 200-PDF set.
| Date | Overall | What changed |
|---|---|---|
| 1 Oct 2026 | 0.648 | First run. Headings came out as bold text, so the heading score was 0. |
| 1 Oct 2026 | 0.788 | Real Word headings that other apps recognise too. |
| 1 Oct 2026 | 0.811 | Tables drawn with only horizontal lines, merged table cells, page headers and footers in the right place. |
| 2 Oct 2026 | 0.825 | Long ruled tables kept whole, row names beside tables kept, contents pages. |
| 11 Oct 2026 | 0.884 | Ten days of real-world fixes: forms, posters, sideways pages, charts, two-column journals and more (see What’s new). |
Altto vs iLovePDF on the hardest PDFs
The other converters in the table are developer tools, not websites, so on 1 October 2026 we also ran the 8 PDFs Altto then found hardest through iLovePDF’s free PDF to Word and scored both the same way. Altto kept tables (0.89 vs 0.76) and headings (0.58 vs 0.37) better; iLovePDF was slightly ahead on reading order (0.90 vs 0.88). The full write-up is in our iLovePDF comparison. We haven’t tested Smallpdf or Adobe Acrobat on this set yet; when we do, the results will go here.
Where Altto still struggles
- Tables that are pictures. Some PDFs contain a table only as an image. Altto keeps the picture, so it looks right but the cells aren’t editable (and score 0 here).
- Designed one-page posters and slides with text dotted around the page. Altto keeps the look, but the benchmark expects one reading order, and there often isn’t an obvious one.
- Charts. Altto keeps charts as pictures, so their labels aren’t editable text. Some converters turn the labels into text, which scores better here even though the chart itself is lost.
Read this before quoting the numbers
- We built and tested Altto on this set. Since 1 October we have used these 200 PDFs to find bugs, so Altto’s score here is probably kinder than its score on PDFs it has never seen. That’s why we also test on our own collection of real-world files (forms, bank statements, posters, journals) before every release.
- Different outputs. Altto makes a Word file; the others make Markdown. To score Altto we turn its Word file into Markdown with pandoc, which can only lose detail, never add it. Pictures are removed before scoring for every tool.
- Not the same date. The other converters were scored by the benchmark’s authors in April 2026 on the versions then current; newer versions may do better.
- Looks aren’t scored. The test checks text, tables and headings, not fonts, colours or page layout, which matter a lot in a Word file.
How to repeat the test
- Download opendataloader-bench (200 PDFs, hand-checked answers and scoring code, Apache-2.0).
- Convert each PDF on the Altto homepage with the default settings. We used headless Chromium to do all 200 automatically; it took about 15 minutes.
- Turn each Word file into Markdown:
pandoc -f docx -t gfm --wrap=none file.docx, and remove picture links. - Put the Markdown files in
prediction/altto/markdownand runpython src/evaluator.py --engine altto.
Common questions
How accurate is Altto's PDF to Word converter?
On the 200 public PDFs of opendataloader-bench, tested on 11 October 2026, Altto scored 0.884 out of 1 overall: 0.906 for reading order, 0.862 for tables and 0.831 for headings. 112 of the 200 documents scored 0.9 or more, and 10 scored under 0.7.
Which converters were compared?
Altto and 14 published results from the benchmark itself: OpenDataLoader (four modes), Nutrient, Docling, Marker, MinerU, Unstructured (two modes), EdgeParse, PyMuPDF4LLM, MarkItDown and LiteParse. Their scores come from the benchmark repository; we did not re-run them. Smallpdf and Adobe Acrobat have not been tested yet.
Can I repeat the test myself?
Yes. The 200 PDFs, the hand-checked answers and the scoring code are public on GitHub (opendataloader-bench, Apache-2.0). Convert each PDF with the Altto homepage, turn the Word files into Markdown with pandoc, and run the benchmark's evaluator. Our per-document scores are in a downloadable JSON file on this page.
Does Altto upload my PDF to run the conversion?
No. The converter tested here runs in your browser on your own device, so the PDF is never uploaded. The test was run the same way, in a browser, with the same page visitors use.
The best test is still your own file. Try the PDF to Word converter: it runs in your browser, so your PDF never leaves your device.