OCRRouter

Marker vs DeepSeek-OCR

Pick Marker for self-hosted PDF to Markdown on books and papers with equations, or when you need RAG chunks, JSON block trees or CPU-only runs. Pick DeepSeek-OCR for cheap bulk OCR on your own CUDA GPUs when you need an MIT licence for commercial use. The main split is licence and hardware: Marker's model weights restrict commercial use for companies above $5M, while DeepSeek-OCR is MIT but needs a CUDA GPU.

Markerdatalab/markerDeepSeek-OCRdeepseek/deepseek-ocr
TypeOCR APIOpen source
Price14 credits / page1 credit / page
Per 1,000 pages$14.00$1.00
Typical speed2.5 s / page1.1 s / page
Best forBooks and research papersCheap bulk conversion
InputsPDF, images, DOCXPDF, images
Languages90+ languages100+ languages
Max pages per request1,000200
Tables✓ Yes✓ Yes
HandwritingNoNo
Math✓ Yes✓ Yes
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
FormsNoNo
Multilingual✓ Yes✓ Yes
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

  • You can run it on CPU, Apple Silicon or GPU, and fast mode needs only about 18 MB of ONNX weights.
  • It outputs Markdown, HTML, JSON block trees, LaTeX and RAG chunks, so it fits books, papers with equations and chunking pipelines.
  • A hosted API from Datalab is available with about 5,000 free pages per month if you do not want to run it yourself.

Choose DeepSeek-OCR if

  • The MIT licence allows commercial use with no revenue cap.
  • It is built for cheap self-hosted bulk OCR, with about 2,500 tokens per second of PDF throughput on one A100-40G (authors' figure).
  • Resolution modes from 64 to 400 vision tokens per page let you trade accuracy for cost, and it can return bounding boxes.

Key differences

MarkerDeepSeek-OCR
LicenceCode Apache-2.0; weights under a modified AI Pubs Open Rail-M licence, free for research, personal use and startups under $5M.MIT
How you run itSelf-host on GPU, CPU or Apple Silicon, or use Datalab's hosted API.Self-host only, with Transformers or vLLM.
HardwareGPU, CPU or MPS. vllm mode needs an NVIDIA GPU. Fast mode runs on CPU.CUDA GPU. Reference setup is CUDA 11.8 and PyTorch 2.6 in bfloat16.
PriceFree self-hosted within the weight licence, you pay for compute. Hosted API has about 5,000 free pages per month, then paid plans.Free self-hosted, you pay for compute.
Output formatsMarkdown, HTML, JSON, chunks, LaTeX, images.Markdown, plain text, HTML, bounding boxes.
BenchmarksolmOCR-bench 83.7 in vllm mode and 66.6 in CPU fast mode. OmniParseBench 85.9 in vllm mode. All are Datalab's own numbers.Not published
Known failureCPU fast mode scores lower (66.6 on olmOCR-bench) than vllm mode (83.7).Repeated tokens and looping over table headers (GitHub issues #12 and #16). A blank image took 30 seconds and produced repeated garbage. It corrects nonsense words instead of copying them, so codes and unusual words may be changed.

Sources

What each one is

Marker

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

Marker details

DeepSeek-OCR

Open-source vision model that compresses pages into few tokens.

DeepSeek-OCR 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": ["deepseek/deepseek-ocr"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Marker alternativesDeepSeek-OCR alternativesBest OCR models in 2026