Why this site exists#
Docling is excellent and moves fast — 233 releases in about two years. That speed makes tutorials go stale: flags get renamed, defaults change, packages split. We built docling.org to be the place where the commands still work, the options are complete and the claims are measured rather than copied.
What we are — and are not#
We are
An independent, community-run reference. We write the guides, build the tools, run the benchmarks and keep them up to date. Contact: hello@docling.org.
We are not
The Docling project, IBM, IBM Research or the LF AI & Data Foundation. We have no access to unreleased plans and do not speak for the maintainers. Official sources: GitHub, docling.ai, official docs.
How we verify#
- Pin a version
Every page states the Docling version it was checked against (currently 2.143.1, verified 27 Oct 2026). When a release changes behaviour, we re-check the affected pages.
- Read options from the software itself
The CLI reference, config builder and engine tables are generated from Docling's own parameter definitions (
typercommand objects,docling options, pydantic model fields), not typed by hand. - Run what we publish
Commands and code snippets are executed in a clean virtual environment. When something does not work — for example a “no-PyTorch” Office install we expected to work — we publish what actually happened.
- Measure, and describe the machine
Benchmarks state hardware, versions, documents and method, and use medians of repeated warm runs. They are small and specific; we say so.
- Cite everything else
Licences, model scores, image sizes and download counts link to where we read them, with the date.
Our benchmark setup#
| Hardware | 4 vCPU x86_64 cloud machine, 15 GB RAM, no GPU |
|---|---|
| Software | Linux, Python 3.13, PyTorch 2.14.1 (CPU build), Docling 2.143.1 |
| Documents | The Docling Technical Report (arXiv 2408.09869, 9 pages) and a rasterised 4-page copy at 150 DPI |
| Other tools | Marker 1.10.2, PyMuPDF4LLM 2.0, MarkItDown 0.1.9 and Unstructured 0.27.35, each in its own virtual environment; Tesseract 5.3.4 for tools that OCR through it |
| Method | One converter per configuration, one warm-up run, median of three timed runs; Docling's built-in stage profiler |
Details and results: performance and Docling vs alternatives.
Corrections and updates#
If a command fails, a number looks wrong or a page is out of date, email hello@docling.org with the page URL, your Docling version (docling --version) and what you saw. We fix confirmed errors quickly and update the “Updated” date on the page.
Languages#
The guide is available in English, 简体中文 and 日本語. Commands and option names are identical in every language; only the explanations are translated.
Money and independence#
docling.org carries no sponsored content and no affiliate links. We use Google Analytics to understand which pages help people (see privacy). Nobody pays us to rank one tool above another; the comparison page shows our raw numbers so you can judge for yourself.
Frequently asked questions
Is docling.org affiliated with IBM?
No. docling.org is independent and is not affiliated with, endorsed by or maintained by IBM, the LF AI & Data Foundation or the Docling project.
How can I report a mistake on docling.org?
Email hello@docling.org with the page address, what is wrong, and your Docling version if a command failed. We aim to publish corrections within a day.
Where is the official Docling documentation?
At docling-project.github.io/docling, with the source code at github.com/docling-project/docling and the project site at docling.ai.