Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
RAG & Document Chat

Marker Review: PDF-to-Markdown Converter with OCR and Table Extraction

·6 min readBy · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Marker (40,000+ GitHub stars) converts PDF, DOCX, PPTX, XLSX, HTML, EPUB and images into Markdown, JSON and HTML with the Surya OCR model. The code is Apache-2.0; the model weights have a separate licence with revenue limits. It is a Python library and CLI; install with pip install marker-pdf.

Key Takeaways

  • Code Apache-2.0, model weights under a separate licence; 40,000+ GitHub stars; developed by Datalab.
  • pip install marker-pdf — Python 3.10+ and PyTorch required.
  • Outputs Markdown, JSON (with layout bounding boxes), HTML and chunked output.
  • Surya VLM handles OCR for scanned PDFs and images; TableConverter handles tables.
  • The README does not mention an MCP server; wrap Marker in your own API if you need AI-assistant access.

📍 In One Sentence

Marker is a Python library (code under Apache-2.0, model weights under a separate licence) that converts PDF, DOCX, PPTX, XLSX, HTML, EPUB and images into Markdown, JSON and HTML using the Surya OCR model, running on your machine or via the Datalab hosted API.

💬 In Plain Terms

You give it a document; it reads the layout and text — including tables and scanned pages — and writes clean Markdown or JSON that a RAG system or LLM can consume. By default Marker processes documents on your machine; the optional --use_llm mode can send content to the LLM provider you configure.

📌Note: PromptQuorum has not tested Marker hands-on. All facts are from github.com/datalab-to/marker and datalab.to, read 2026-11-15.

Marker is an open-source Python library and CLI that converts PDF, DOCX, PPTX, XLSX, HTML, EPUB and image files into Markdown, JSON and HTML, using the Surya vision-language model for OCR. It is developed by Datalab, which also offers a hosted API at datalab.to with a monthly free allowance. This review is based on the Marker GitHub repository and datalab.to, checked on 15 November 2026; PromptQuorum has not installed or run it.

Marker Review: PDF-to-Markdown Converter with OCR and Table Extraction

What Marker Does

  • Input formats: PDF, PPTX, DOCX, XLSX, HTML, EPUB, and common image formats (per README, 2026-11-15).
  • Output formats: Markdown (default), JSON with layout bounding boxes, HTML, and "chunks" for RAG.
  • OCR: Surya VLM — a vision-language model that handles multi-language scanned documents.
  • Table extraction: TableConverter re-draws tables in Markdown; structured JSON output preserves row and column data.
  • Bounding boxes: JSON output includes per-block layout coordinates, useful for citation mapping and UI overlays.
  • GPU optional: Surya works on GPU, CPU or Apple MPS; GPU is faster but not required.
  • Hosted API (Datalab): monthly free allowance (the README cites about 5,000 convert pages; see datalab.to/pricing); paid tiers for higher volume.

Install Marker

  • Local (Python): pip install marker-pdf for PDF; `pip install marker-pdf[full]` for other formats (PPTX, DOCX, XLSX, HTML, EPUB).
  • PyTorch: install separately first if not already present (pip install torch); GPU build is optional.
  • Surya server: on first use Surya starts a local inference server; it needs Docker with vllm and the NVIDIA Container Toolkit on NVIDIA GPUs, or the llama.cpp llama-server binary on CPU and Apple Silicon.
  • CLI: marker_single input.pdf --output_dir output_dir/ for a single file; marker input_dir/ --output_dir output_dir/ for batch.
  • Hosted API: create an account at datalab.to; monthly free allowance (see datalab.to/pricing); no local Python needed.
  • Source: github.com/datalab-to/marker (code: Apache-2.0).

Data Flow

  • By default Marker processes documents on your machine. The optional --use_llm mode can send content to the LLM provider you configure (default gemini-3.5-flash; Claude, OpenAI-compatible, Azure, Vertex, OpenRouter and Ollama are also supported). The README does not say when model weights are downloaded; expect a download during setup.
  • If you use the Datalab hosted API, your documents are uploaded to Datalab servers.
  • This review did not assess Datalab's privacy terms; check them on datalab.to before sending confidential documents.
  • For sensitive workloads, run Marker locally without --use_llm.

Licence and Price

  • Licence: the code is Apache-2.0; the model weights use a modified AI Pubs Open RAIL-M licence, free for research and personal use and for organisations with no more than USD 5 million gross revenue in the prior year and no more than USD 5 million total funding raised. Larger organisations need a commercial licence from Datalab, and the licence bars use in products or services that compete with Datalab.
  • Local use: no account needed; free within the weights-licence terms above.
  • Datalab API (hosted): monthly free allowance (the README cites about 5,000 convert pages; the pricing page states it as a dollar amount: $20/month on a work email, $10 on a personal one); see datalab.to/pricing for current terms.
  • Check the model-weights licence before any commercial use; the Apache-2.0 licence covers only the code.

📌Note: Warning: check the model licence before commercial use. Only the Marker code is Apache-2.0. The model weights use a modified AI Pubs Open RAIL-M licence; organisations above USD 5 million gross revenue or USD 5 million total funding need a commercial licence from Datalab.

Competitors and Alternatives

  • Docling — IBM MIT-licensed parser; broader output formats (DocLang, DocTags) and a documented MCP server (the separate docling-mcp package); also local.
  • MinerU — open-source; outputs Markdown, HTML and LaTeX; cross-page table merging; MinerU Open Source License (Apache 2.0 plus additional conditions; check before commercial use).
  • Unstructured — Python ETL library; broad format support including emails; outputs JSON elements; typically needs Tesseract and Poppler for local OCR and PDF handling.
  • OpenDataLoader PDF — PDF-only parser with Python, Node.js and Java SDKs; outputs JSON with bounding boxes, Markdown and HTML.

Compare in directory

Frequently Asked Questions

Is Marker free to use commercially?

It depends on your organisation. The code is Apache-2.0, but the model weights use a modified AI Pubs Open RAIL-M licence: free for research and personal use, and for organisations with no more than USD 5 million gross revenue in the prior year and no more than USD 5 million total funding raised. Larger organisations need a commercial licence from Datalab, and the licence bars use in products or services that compete with Datalab. Check the model licence before commercial use.

What is the difference between Marker and Docling?

Both convert PDFs to Markdown and JSON locally. Marker uses the Surya OCR model and focuses on clean Markdown output; Docling (MIT, IBM) has a documented MCP server (separate docling-mcp package) so AI assistants can call it directly, and outputs additional formats (DocLang, DocTags).

Does Marker need a GPU?

No. Surya VLM runs on CPU. The README states that Marker works on GPU, CPU or MPS; a GPU speeds up processing but is not required.

How do I install Marker?

Run pip install marker-pdf in a Python 3.10+ environment. Install PyTorch separately if needed. For formats other than PDF (PPTX, DOCX, XLSX, HTML, EPUB) use `pip install marker-pdf[full]`. On first use Surya starts a local inference server, which needs Docker with vllm and the NVIDIA Container Toolkit on NVIDIA GPUs, or the llama.cpp llama-server binary on CPU or Apple Silicon. No GUI installer or desktop app exists. The Datalab hosted API requires no local installation.

Can Marker handle scanned PDFs?

Yes. Marker uses the Surya vision-language model for OCR, which handles scanned documents and image-only PDFs.

What output formats does Marker produce?

Markdown (default), JSON (with per-block bounding boxes), HTML and chunked output for RAG. Markdown is the primary target; JSON is useful when you need layout coordinates.

Does Marker have an MCP server?

The Marker README does not mention an MCP server, so none is documented. If you need a documented AI-assistant integration, consider Docling, which lists one.

What is the Datalab API?

Datalab develops Marker and offers a hosted API. It processes documents in the cloud and has a monthly free allowance (the README cites about 5,000 convert pages; the pricing page states it as a dollar amount: $20/month on a work email, $10 on a personal one); see datalab.to/pricing for current terms. It is useful if you do not want to manage a local Python environment.

Does Marker extract tables?

Yes. Marker uses a TableConverter component that reconstructs tables in Markdown and JSON output. The README documents this as a supported feature.

What version is current?

v2.2.0, released 2026-11-12 (GitHub Releases). PromptQuorum verified this on 2026-11-15.

Sources

  • github.com/datalab-to/marker — README and release history, read 2026-11-15.
  • datalab.to — hosted API information, read 2026-11-15.
  • GitHub Releases: v2.2.0 published 2026-11-12; repository last pushed 2026-11-14.

← Back to Power Local LLM

MCP