OCRRouter

Marker vs Granite Docling

Pick Marker for self-hosted PDF to Markdown on books, papers and RAG pipelines, if your company is under $5M or you buy a Datalab license. Pick Granite Docling for English PDFs with tables, code and equations inside a Docling pipeline, mainly because it is Apache-2.0 with no revenue limit and only 258M parameters. Check Granite Docling on invoices and page headings first, since it can get stuck in repetition loops there.

Markerdatalab/markerGranite Doclingibm/granite-docling
TypeOCR APIOpen source
Price14 credits / page1 credit / page
Per 1,000 pages$14.00$1.00
Typical speed2.5 s / page1.5 s / page
Best forBooks and research papersSmall-model document conversion
InputsPDF, images, DOCXPDF, images
Languages90+ languagesEnglish, experimental CJK and Arabic
Max pages per request1,000500
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ Yes✓ Yes
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ YesNo
On OCRRouterAvailableAvailable

Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.

Which to choose

Choose Marker if

  • It has a CPU fast mode with about 18 MB of ONNX weights that runs at 23.6 pages per second, and a vllm mode for NVIDIA GPUs.
  • It outputs Markdown, HTML, JSON block trees, RAG chunks, LaTeX and images, so one tool covers chunking for retrieval.
  • Datalab also sells a hosted API with about 5,000 free pages per month, so you can skip running it yourself.

Choose Granite Docling if

  • The whole package is Apache-2.0, so large companies can use it commercially without a paid license.
  • At 258M parameters it suits low-cost CPU or small GPU conversion, and it reports table structure TEDS of 0.97 on FinTabNet.
  • It returns bounding boxes and DocTags along with Markdown, HTML and LaTeX, and it fits directly into Docling pipelines.

Key differences

MarkerGranite Docling
LicenceCode Apache-2.0. Weights under a modified AI Pubs Open Rail-M license, free for research, personal use and startups under $5M.Apache-2.0
Commercial useCompanies above $5M need a paid license for the weights.No stated restriction beyond Apache-2.0.
How you run itSelf-host on GPU, CPU or Apple Silicon (MPS), or use Datalab's hosted API. vllm mode needs an NVIDIA GPU.Self-host only, typically inside Docling. Model size is 258M parameters.
Output formatsMarkdown, HTML, JSON, chunks, LaTeX, images.DocTags, Markdown, HTML, LaTeX, bounding boxes.
Benchmarks (published by the authors)olmOCR-bench 83.7 in vllm mode, 66.6 in CPU fast mode. OmniParseBench 85.9 in vllm mode.FinTabNet table TEDS structure 0.97, full-page OCR F1 0.84, equation F1 0.968, OCRBench 500.
LanguagesNot publishedEnglish primary. Japanese, Arabic and Chinese are marked experimental.
Known failure modesCPU fast mode scores lower on olmOCR-bench (66.6 versus 83.7 in vllm mode).Repetitive loops on page headings and invoices. May output only exclamation marks on GPUs without bfloat16, such as T4, unless run in float32. vLLM can fail with a wte attribute error and needs the untied revision.

Sources

What each one is

Marker

Hosted version of the open-source Marker and Surya models. Strong on books and papers.

Marker details

Granite Docling

Compact IBM vision model that converts pages to DocTags for the Docling pipeline.

Granite Docling details

Try both on your own file

Run the file once with each model and compare the output.

PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?

Use one, fall back to the other

{"model": "datalab/marker","fallbacks": ["ibm/granite-docling"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Marker alternativesGranite Docling alternativesBest OCR models in 2026