Skip to content
dialegein

Docling

IBM-started open-source library that turns PDFs and office documents into structured Markdown, HTML or JSON with layout, table and reading-order models.

Jobs Extract Convert OCR

Since 2024

Operated by LF AI & Data Foundation

Visit Docling

Opens the vendor's own site.

Jobs covered

Open a task to see exactly what the vendor documentation supports.

Extract Browse all extract tools

extracting structured content from documents with the Docling Python library or CLI, which parses formats including PDF, DOCX, PPTX, XLSX, HTML, EPUB and images into the unified DoclingDocument representation and exports it to Markdown, HTML, JSON, DocTags or DocLang, with the advanced PDF understanding the project lists as page layout, reading order, table structure, code and formulas

Convert Browse all convert tools

renditions only: exporting the parsed Docling document to serialized files with the library export methods, Markdown through export_to_markdown, HTML through export_to_html, plain text through export_to_text, DocTags and DocLang, VTT and a dictionary form, which serialise the Docling Document rather than converting between document formats, with no DOCX, PPTX, XLSX or PDF output, no round trip and no fidelity claim

OCR Browse all ocr tools

running OCR on scanned pages and images in Docling with a selectable engine among RapidOCR, Nemotron-OCR, EasyOCR, ocrmac, tesseract CLI and tesserocr, each installed as an extra package, with the language set through the single OcrOptions.lang field either as engine native codes or as iso: prefixed BCP-47 tags, an error raised when the selected engine has no model for the requested language, and an option to send OCR requests to a remote KServe v2 inference server

Where your files are processed

No server involved

An in-process Python library and CLI with no hosted component and no account. One exception, recorded in the note: the optional VLM picture-description path and the convert-remote command can be pointed at a server, so page images leave only if the operator configures it that way. That is a property of the library itself: what an application built with it does with your file depends on that application, not on this.

Architectural: pip install docling gives an in-process library plus a local CLI, with no hosted component in the package and no account. The project's README lists local execution for sensitive data and air-gapped environments as a feature.

One exception to be aware of: the same package can be pointed at an OpenAI-compatible inference server for VLM and picture-description enrichment, and `docling convert-remote` talks to a docling-serve instance; in those configurations page images do leave the machine, by the operator's own choice.

What it does

Docling converts PDF, DOCX, PPTX, XLSX, HTML, EPUB, LaTeX, email, audio and image files into a single in-memory document model described by layout, reading order, table structure and picture classification, then exports that model as Markdown, HTML, JSON, DocTags, WebVTT or DocLang. It runs entirely in-process as a Python library with a CLI and a bundled HTTP server, and it is the one of the two rivals here that can be driven with no GPU at all on a CPU-only install.

Layout, table and OCR models are downloaded from Hugging Face on first use.

Best for

Teams that need a permissively licensed converter they can embed in a Python pipeline and run on their own machines, including air-gapped and CPU-only ones, and who care about breadth of input formats more than top-end accuracy on ugly scans.

Worth knowing

Accuracy on hard PDFs is its weak point, and the proof is the competitor's own benchmark: Datalab's olmocr-bench comparison scores Docling at 50.3 overall and 64.0 on born-digital PDFs against Marker's 76.0 and 83.5. On top of that, the first conversion downloads several gigabytes of model weights, so a cold start on a locked-down or air-gapped machine fails unless the models are pre-fetched, and the default OCR engine selection is a real quality decision rather than a default you can ignore.

Details, and where each comes from

Vendor statements, our own reading of their pages, and the things nobody has established yet, each answered separately. Where we have not confirmed something it says so rather than guessing, and a row nobody has established is greyed out.

Versions covered
Windows · macOS · Linux · API · Self-hosted
Website
docling-project.github.io
Requires an account
Not applicable
Free tier
Yes
Watermark on the free tier
Not verified
Maximum file size
Not verified
Batch processing
Yes
Data residency
Self-hosted
Open source
Yes
Self-hostable
Yes
Licence
MIT for the Docling codebase.

MIT for the Docling codebase. The optional OCR and vision-language model weights are licensed separately by their original packages, and the README directs users to those licences rather than covering them.

Redaction
No redaction feature
Programmatic access
Embeddable library · Command line · HTTP API
Last availability check
Responded in 96 ms 21 October 2026

Checked automatically by our own runner, not reported by the vendor. 100% of our last 30 checks got a response.

What the redaction verdict rests on

From the vendor's own documentation, read 22 September 2026. Not tested by us.

What that does not establish

The vendor’s own site was read on 22 September 2026 and contains no mention of redaction, sanitisation, blackout or whiteout. Absence from a vendor’s own material is not proof that a feature does not exist - it may be undocumented, or on a page the reading did not reach - but for a NEGATIVE claim it is the strongest evidence available, and it is weaker than a positive quote.

Recorded as such rather than presented as certainty.

https://docling-project.github.io/docling/

What the vendor says

Claims are labelled rather than repeated as fact. Where we have checked something ourselves, it says so and shows when. Checked means we read the vendor's own page and it says what we recorded - not that we tested the behaviour, unless the note says we did.

  • “Advanced PDF understanding incl. page layout, reading order, table structure, code, formulas, image classification, and more”

    Checked by us - the source says this Read 22 September 2026

    Confirmed verbatim: the line appears unaltered in the README's feature list, with its emoji prefix stripped in the rendered text. It is a capability list rather than a measured claim, and it is worth reading beside the neighbouring Docling entry in this index - IBM's hosted service claiming "~ 14x faster document processing compared to open-source Docling" - because the two together describe the open-source project and the commercial service built on it.

    The project's own README makes no speed claim against anything. Recorded as the project's description; no library was installed.

    Standing: Re-read and matched byte for byte

    Check it yourself: the source, the SHA-256, and how we checked
    Read
    https://github.com/docling-project/docling
    From
    the text a browser renders
    SHA-256
    8bcd2e8459e3164aa60d9459cae87fb6294c5b9dbe51fca9cf62801aff11adf1

    How we checked

    Re-verified verbatim on 2026-09-22: https://github.com/docling-project/docling re-rendered in headless Chrome with scripts/render-page.mjs and the claim's wording found in the rendered text. Originally verified on 2026-09-21. sourceHash is the SHA-256 of the re-rendered text, held as rv-github-com-docling-project-docling-9f316614.txt.

    What this hash proves, and what it does not

  • “~ 14x faster document processing compared to open-source Docling”

    Checked by us - the source says this Read 21 September 2026

    Confirmed verbatim, and the method is part of the finding: a reader of the rendered page sees a number that starts at zero and counts up, so the claim is only checkable against the markup. What the page does not carry is any basis for the figure.

    There is no benchmark, no hardware, no corpus and no date attached to it, and the comparison is IBM's hosted service against the open-source Docling project that IBM also publishes - the same project against which it is being measured. Two neighbouring statistics on the same page behave identically (downloads shown as "0M+", price as "USD 0") and are animated in the same way.

    So this is a vendor marketing figure that cannot be checked from the page, recorded as the vendor's claim and not as a measurement.

    Standing: Re-read and matched once per-response tokens were masked

    Check it yourself: the source, the SHA-256, and how we checked
    Read
    https://www.ibm.com/products/docling
    From
    the raw HTTP response
    SHA-256
    983c4cab1ddd483bd59706dcba96c7457b0fa54a8b93744061f2f7e05b4e7cd4

    How we checked

    The figure is not in the rendered text: headless Chrome produced "~ 0x" where the claim has "~ 14x", because the statistic is a JavaScript counter that animates up from zero and the render captured it before it ran. The claim was therefore read from the raw HTML response, 266117 bytes saved and hashed, where the statistic block contains "~ 14x" followed by "faster document processing compared to open-source Docling". sourceHash is the SHA-256 of that raw response.

    What this hash proves, and what it does not

Each claim is quoted from the vendor's own pages. This section describes what we have and have not checked - it is not a judgement about anyone.

Sponsorship is distribution, not an endorsement or a promise of search rankings, AI citations, traffic, or sales. This listing is a free entry, compiled from published sources.

Visit Docling Report a problem

Related reading

How this record was put together

Editorial review.
No editor has signed this record off. That is not the same as nothing having been checked, and the lines below say what has.
Source reading.
2 claims taken from the vendor's own pages, read between 21 September 2026 and 22 September 2026.
Website availability.
The vendor's site was last checked on 21 October 2026.
Behaviour testing.
We ran this ourselves. 1 claim here rest on something we observed rather than something a vendor wrote. The file handling was watched directly.

Sources: docling-project.github.io/docling (usage index, OCR concepts page, API server managed-service page); github.com/docling-project/docling (README, LICENSE, pyproject.toml); raw docs/usage/gpu.md, docs/getting_started/installation.md and docs/examples/pii_obfuscate.py from the same repository; pypi.org/pypi/docling JSON API for release dates; api.github.com repository metadata; ibm.com/products/docling product page, re-read 2026-09-25 for the managed-service hosting line ('Managed API-based Docling service. Instance deployed on AWS.'); github.com/datalab-to/marker README, read 2026-09-25, for the olmocr-bench composition and its own wording that olmocr-bench is 'a third-party benchmark' linking github.com/allenai/olmocr.