Qianfan OCR vs SmolDocling
Pick Qianfan OCR for self-hosted parsing of multilingual documents, handwriting, tables and formulas, where it posts 93.12 on OmniDocBench v1.5. Pick SmolDocling only for English page conversion experiments on small hardware, since it is a 256M preview model that its authors say is replaced by Granite-Docling.
Qianfan OCRbaidu/qianfan-ocr | SmolDoclinghuggingface/smoldocling | |
|---|---|---|
| Type | OCR API | Open source |
| Price | 2 credits / page | 1 credit / page |
| Per 1,000 pages | $2.00 | $1.00 |
| Typical speed | 2.0 s / page | 1.0 s / page |
| Best for | Chinese documents | Very cheap page conversion |
| Inputs | PDF, images | PDF, images |
| Languages | Chinese, English and more | Mainly English |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | No | ✓ Yes |
| Confidence | No | 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 Qianfan OCR if
- You have documents in languages other than English, because it covers 192 languages plus handwriting and scene text.
- You need key information extraction, tables or formulas, with output as Markdown, JSON, HTML, LaTeX or bounding boxes.
- You want a model with published scores of 93.12 on OmniDocBench v1.5 and 79.8 on olmOCR-Bench, under Apache-2.0 for commercial use.
Choose SmolDocling if
- You only process English pages and want about 0.35 seconds per page on an A100 with vLLM.
- You have modest hardware, since the model has only 256M parameters.
- You want to experiment with a tiny document VLM that outputs DocTags for Docling and also converts to Markdown, HTML and JSON.
Key differences
| Qianfan OCR | SmolDocling | |
|---|---|---|
| Size | 4B parameters (Qwen3-4B language model) | 256M parameters |
| Languages | 192 languages, plus handwriting and scene text | English only |
| OCRBench | 880 | 338 |
| Speed | 1.024 pages per second on an A100 with W8A8; thinking mode can emit up to 16,384 tokens | About 0.35 seconds per page on an A100 with vLLM |
| How you run it | Self-host the open weights (benchmarked on one A100 with vLLM) or use the Baidu Qianfan platform | Self-host only |
| Licence | Apache-2.0 | Apache-2.0 per model summary; metadata tag says CDLA-Permissive-2.0 |
| Known failure | Thinking mode adds latency; not suited to CPU-only deployment | Endless token repetition and missing closing tags; output loops on some images (Docling issue #1515), making generation very slow |
Sources
- Qianfan-OCR model card
- Qianfan-OCR paper
- SmolDocling model card
- SmolDocling paper
- Docling issue #1515
- Granite-Docling comparison table
What each one is
Qianfan OCR
Baidu's OCR-tuned vision model. Strong on Chinese documents and tables.
Qianfan 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.
PDF, PNG, JPG, WebP, HEIC. Up to 20 MB each, 10 files.No file?
Use one, fall back to the other
{"model": "baidu/qianfan-ocr","fallbacks": ["huggingface/smoldocling"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qianfan OCR vs MonkeyOCRQianfan OCR vs Granite DoclingPaddleOCR-VL vs SmolDoclingJina OCR vs SmolDoclingSmolDocling vs Phi-4-multimodalSmolDocling vs Google Cloud Vision OCRKraken vs SmolDoclingQianfan OCR vs Chandra
Qianfan OCR alternativesSmolDocling alternativesBest OCR models in 2026