IBM Docling: Open-Source Document Conversion for AI
Docling is an open-source toolkit that turns documents into structured output for AI applications. IBM Research Zurich started it in July 2024. It now lives at the LF AI & Data Foundation under the docling-project organization on GitHub. You give it a PDF or a scanned page. It gives back Markdown or lossless JSON, with headings and tables intact. Docling keeps the tables and headings that plain text extraction loses. That matters when PDFs feed a retrieval pipeline or an agent. This guide shows how it works and how to install it, then checks which speed claims hold up.
What is Docling, in one sentence?
Docling is an MIT-licensed Python library and CLI that parses documents into a structured format you can export as Markdown or JSON for retrieval and agents.
The official GitHub repository describes its purpose as "Get your documents ready for gen AI." It had 69,776 stars on October 28, 2026, per the GitHub API. Releases are frequent: the latest was v2.144.0, published the same day. The README says the code is MIT-licensed and each bundled model has its own license.
IBM described the project's origin in a November 2024 research blog post. In April 2025, IBM contributed Docling to the Linux Foundation's LF AI & Data Foundation, per the Linux Foundation announcement.
How does Docling differ from plain OCR or text extraction?
Most PDF tools either pull the embedded text layer or run OCR on page images. Docling runs layout and table models first, so it knows which text is a heading or a table cell.
| Feature | Docling | Text-layer extraction (e.g. pdftotext) | OCR only (e.g. Tesseract) |
|---|---|---|---|
| Reading order and headings | Recovered by a layout model | Often lost on multi-column pages | Not recovered |
| Tables | Rebuilt as rows and cells by TableFormer | Flattened into text | Flattened into text |
| Scanned pages | OCR when needed | Fails (no text layer) | Works |
| Output | Markdown, HTML, lossless JSON, DocTags | Plain text | Plain text |
| Runs locally | Yes, including air-gapped setups | Yes | Yes |
Sources: the Docling README and supported formats page. Model details are in the Docling technical report.
What formats does Docling support?
The supported formats page lists far more than PDF.
| Group | Examples | Notes |
|---|---|---|
| PDF and images | PDF, PNG, JPEG, TIFF, WEBP | OCR runs where there is no text layer |
| Office | DOCX, XLSX, PPTX | Older DOC, XLS and PPT need LibreOffice |
| Markup and text | HTML, Markdown, AsciiDoc, LaTeX, CSV | |
| Publishing and mail | EPUB, EML, MSG, JATS and USPTO XML | |
| Audio and video | WAV, MP3, MP4 | Needs the optional ASR extra |
The usual output is Markdown or HTML. For lossless round trips, Docling writes JSON of its DoclingDocument model. DocTags and chunked JSONL are also available.
How does Docling work?
The standard pipeline reads a page in stages. A layout model finds the blocks on the page, such as headings and tables. The 2025 technical report says its architecture "is derived from RT-DETR" and was trained on DocLayNet plus proprietary data. TableFormer then rebuilds each table's grid, including merged cells. OCR runs only where the page has no usable text.
The OCR engine is pluggable. The installation docs name RapidOCR as the default. EasyOCR and Tesseract are optional, along with OcrMac on Apple machines.
There is also a vision-language pipeline. Here one small VLM reads the page image and writes DocTags. The current model is granite-docling-258M, released under Apache 2.0 in September 2025. Its model card calls it the successor to SmolDocling.
How do you install and use Docling?
Install it with pip. The README requires Python 3.10 or newer and supports all three desktop operating systems:
pip install docling
The CLI converts a file or URL to Markdown in the current directory:
docling https://arxiv.org/pdf/2206.01062
In Python, the README shows a DocumentConverter. You pass it a path or URL and export the result:
from docling.document_converter import DocumentConverter
converter = DocumentConverter()
result = converter.convert("https://arxiv.org/pdf/2408.09869")
print(result.document.export_to_markdown())
For the VLM pipeline, the granite-docling model card gives the command docling --pipeline vlm --vlm-model granite_docling <url>. For a service, docling-serve wraps the converter in an HTTP API with an optional web UI.
How do AI agents use Docling?
Two routes are common. Retrieval frameworks call it as a loader. The Docling docs list integrations with LangChain, LlamaIndex, Haystack and CrewAI.
Coding and desktop agents use the Docling MCP server. Its README starts it with uvx --from docling-mcp docling-mcp-server --transport stdio. The tools convert documents to DoclingDocument JSON and export them to Markdown. Others let an agent build a new document step by step. Any MCP host can call it, so Claude Code or Cursor gets the same layout-aware conversion as your retrieval pipeline.
What should you check in Docling's performance claims?
Older coverage, including this page before its rewrite, repeated that Docling is "30x faster than OCR." No benchmark backs that wording. The figure comes from a quote in IBM's 2024 blog: a researcher said avoiding OCR speeds up time-to-solution by 30 times. Neither technical report contains a 30x result.
The reports do publish measured numbers. With OCR off, the first report converted 225 pages on an M3 Max laptop at 1.27 to 2.45 pages per second. Thread count and PDF backend explain the spread. The 2025 report measured 3.1 seconds per page on an x86 CPU and 0.49 seconds per page on an Nvidia L4 GPU. On that GPU, MinerU was faster at 0.21 seconds per page. These are the Docling team's own tests on specific hardware. Scanned pages that need OCR will be slower. None of these tests used your documents, so run Docling on a sample of your own files.
The bottom line
Use Docling when your documents have tables or scans and you need their structure. Start with the default pipeline and Markdown export. Test it on 20 of your hardest files before you build on it. Try the granite-docling VLM pipeline if the standard models struggle with your layouts. Give agents the MCP server so they convert files the same way your retrieval pipeline does.
How do agents remember what they learned from documents?
An agent that works on your documents forgets its own conclusions when the session ends. It may learn which pipeline settings fix a vendor's invoice tables, or which sections of a 300-page spec answer recurring questions.
memU keeps those lessons as readable Markdown. A scheduled background task slices new session logs into jobs. The agent distills each job into memory and skill Markdown, then commits it. memU's MemoryService makes no LLM or chat calls. It stores and embeds what the agent wrote, then retrieves it.
A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. A conversion workflow that worked once can become a reusable skill. All hosts share one backend, so a lesson learned in Claude Code is available to Codex.
To try the hosted version, which is free and syncs across devices, get an API key from memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU
memU is open source under Apache-2.0 on GitHub, where the support matrix lists each host.
FAQ
Is Docling made by IBM?
IBM Research Zurich started Docling and open-sourced it in July 2024. IBM contributed it to the LF AI & Data Foundation in April 2025. It is now developed under the docling-project organization on GitHub.
Is Docling free for commercial use?
The Docling code is under the MIT license. Each model has its own license, and granite-docling-258M is Apache 2.0, so check the model you use.
Does Docling need a GPU?
No. Docling runs on CPU, and its technical report measured about 3.1 seconds per page on an x86 CPU. A GPU speeds up the model stages.
Where is the official Docling GitHub repository?
It is at github.com/docling-project/docling. The MCP server and docling-serve live in separate repositories in the same organization.
Can Docling convert scanned PDFs?
Yes. Docling runs OCR on pages without a usable text layer, with RapidOCR as the default engine and EasyOCR or Tesseract as options.