PaddleOCR (PP-OCR) vs SmolDocling
Pick PaddleOCR (PP-OCR) for production text recognition on printed Chinese, English and other Latin-script text, especially on CPU, edge or in the browser. Pick SmolDocling only to experiment with converting English pages to Markdown or DocTags, since it is a preview release that its authors have superseded with Granite-Docling. The main split is output: PP-OCR returns text lines and boxes, SmolDocling returns structured page markup with tables and formulas.
PaddleOCR (PP-OCR)paddle/paddleocr | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| Type | Open source | Open source |
| Price | 1 credit / page | 1 credit / page |
| Per 1,000 pages | $1.00 | $1.00 |
| Typical speed | 0.7 s / page | 1.0 s / page |
| Best for | Fast printed text, Chinese and English | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | 80+ languages | Mainly English |
| Max pages per request | 500 | 500 |
| Tables | No | ✓ Yes |
| Handwriting | No | No |
| Math | No | ✓ Yes |
| Bounding boxes | ✓ Yes | ✓ Yes |
| Confidence | ✓ Yes | No |
| Forms | No | No |
| Multilingual | ✓ Yes | No |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose PaddleOCR (PP-OCR) if
- It covers 50 languages in one PP-OCRv6 model, including Chinese, English, Japanese and 46 Latin-script languages, while SmolDocling is English only.
- The PP-OCRv6 medium model is 34.5M parameters and reports 0.13s on A100, with a 5.2x CPU speedup using OpenVINO.
- It runs from the paddleocr Python package, C++ or the PaddleOCR.js browser SDK, and supports CPU, NVIDIA GPU and Kunlunxin XPU.
Choose SmolDocling if
- It outputs DocTags, Markdown, HTML, JSON and LaTeX in one pass, so you get page structure without a second pipeline.
- One 256M model handles text, tables (OTSL), code, formulas and charts with location tags.
- It reports about 0.35 seconds per page on an A100 with vLLM, and the authors score table structure at TEDS 0.82 on FinTabNet 150dpi.
Key differences
| PaddleOCR (PP-OCR) | SmolDocling | |
|---|---|---|
| Licence | Apache-2.0 | Apache-2.0 per the model summary; the metadata tag says CDLA-Permissive-2.0 |
| Output | Plain text, bounding boxes and JSON. Tables and formulas need the separate PP-Structure or VL pipelines. | DocTags, Markdown, HTML, JSON, LaTeX and bounding boxes |
| Model size | PP-OCRv6 medium: 34.5M parameters | 256M parameters |
| Languages | 50 languages in one PP-OCRv6 model | English only |
| Speed | 0.13s on A100; 5.2x CPU speedup with OpenVINO | About 0.35 seconds per page on an A100 with vLLM |
| Benchmarks | Not published | Table TEDS 0.82 (FinTabNet 150dpi), full-page OCR F1 0.80, OCRBench 338, all by the authors |
| Known failure | CPU inference 4 to 11x slower from denormals in PP-OCRv6 weights (RapidOCR issue #743); NotImplementedError with PaddleOCR 3.7.0 on PaddlePaddle 3.3.1 (issue #18162) | Endless token repetition and missing closing tags; output loops on some images and generation becomes very slow (Docling issue #1515) |
Sources
- PaddleOCR GitHub README
- RapidOCR issue #743 denormal weights
- SmolDocling model card
- SmolDocling paper
- Docling issue #1515
- Granite-Docling comparison table
What each one is
PaddleOCR (PP-OCR)
Baidu's classic open-source text detection and recognition pipeline. Light enough to run on CPU.
PaddleOCR (PP-OCR) detailsSmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "paddle/paddleocr","fallbacks": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
PaddleOCR (PP-OCR) alternativesSmolDocling alternativesBest OCR models in 2026