OCRRouter

PaddleOCR-VL vs docTR

Pick PaddleOCR-VL when you need full pages turned into Markdown or JSON, with tables, formulas, charts and many scripts, and you have a GPU. Pick docTR when you need word-level text with coordinates from invoices or receipts, you want to run on CPU, or you plan to train your own models. The main split is page parsing versus a text detection and recognition library.

PaddleOCR-VLpaddle/paddleocr-vldocTRmindee/doctr
TypeOpen sourceOpen source
Price6 credits / page1 credit / page
Per 1,000 pages$6.00$1.00
Typical speed1.3 s / page0.9 s / page
Best forCJK documentsWord-level OCR with positions
InputsPDF, imagesPDF, images
Languages109 languagesMainly Latin scripts
Max pages per request500500
Tables✓ YesNo
HandwritingNoNo
Math✓ YesNo
Bounding boxes✓ Yes✓ Yes
Confidence✓ Yes✓ Yes
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 PaddleOCR-VL if

  • It outputs Markdown and JSON for full pages and recognizes text, tables, formulas and charts.
  • It supports 109 languages, including Cyrillic, Arabic, Devanagari and Thai, in one model.
  • Version 1.6 scores 96.33 on OmniDocBench v1.6, as reported by its authors, and the model is under 1B parameters, so it fits much smaller GPUs than 7B document VLMs.

Choose docTR if

  • It returns a nested page, block, line and word structure with coordinates as JSON, which suits custom extraction from invoices and receipts.
  • It gives you a choice of detection and recognition architectures, with straight or rotated boxes, and it fits Python ML teams training their own models.
  • It runs on CPU or NVIDIA GPU and ships ready Docker images, a demo on Hugging Face Spaces and a key information extraction predictor.

Key differences

PaddleOCR-VLdocTR
What it isVision-language document parser built on a NaViT-style encoder and ERNIE-4.5-0.3B, 0.9B parametersTwo-stage text detection and recognition library from Mindee, runs on PyTorch
Licence and priceApache-2.0, free, self-hosted (you pay for compute)Apache-2.0, free, self-hosted (you pay for compute)
OutputMarkdown and JSON, including tables, formulas and chartsJSON, plain text and bounding boxes. No table, formula or Markdown output
HardwareNVIDIA GPU recommended, CPU possible but slow, no minimum VRAM statedCPU or NVIDIA GPU, GPU Docker images need CUDA 12.2 or newer on the host
BenchmarksOmniDocBench v1.6 96.33 (authors), olmOCR-Bench 80.0 (independent)Not published
Languages109 languages in one modelNot published
Known failuresVersion 1.6 regresses Russian and Kazakh OCR on born-digital PDFs, decodes about 11x slower because use_cache ships as false, can drop table content, and hangs on blurry imagesSplit-crop merging misjoins readings across repeated characters, cropped images give poorer results, and ONNX export of MASTER truncates decoding steps

Sources

What each one is

PaddleOCR-VL

Open-source document model from Baidu. Excellent on Chinese, Japanese and Korean.

PaddleOCR-VL details

docTR

Open-source OCR library from Mindee with separate text detection and recognition models.

docTR 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": "paddle/paddleocr-vl","fallbacks": ["mindee/doctr"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

PaddleOCR-VL alternativesdocTR alternativesBest OCR models in 2026