Qianfan OCR vs Surya
Pick Qianfan OCR if you self-host on a GPU and need handwriting, scene text, key information extraction or Markdown output across 192 languages. Pick Surya if you need layout detection, reading order and bounding boxes, or want to run locally on a laptop or CPU. The main split is scope: Qianfan OCR is a 4B end-to-end parser, while Surya is a 689M toolkit with a lighter footprint and a revenue-based licence limit.
Qianfan OCRbaidu/qianfan-ocr | Suryadatalab/surya | |
|---|---|---|
| Type | OCR API | Open source |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 2.0 s / page | 1.5 s / page |
| Best for | Chinese documents | Layout and line-level OCR |
| Inputs | PDF, images | PDF, images |
| Languages | Chinese, English and more | 90+ languages |
| Max pages per request | 200 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | ✓ Yes | No |
| Math | ✓ Yes | No |
| Bounding boxes | No | ✓ Yes |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | ✓ Yes |
| 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
- It covers 192 languages, handwriting and scene text, which Surya lists as weak spots or not-for-use cases.
- Apache-2.0 applies to the weights, so there is no revenue or funding cap on commercial use.
- It scores 93.12 on OmniDocBench v1.5 and outputs Markdown as well as JSON, HTML, LaTeX and bounding boxes.
Choose Surya if
- It runs on CPU and Apple Silicon through llama.cpp, with peak memory of about 4.3 to 5.9 GB on an M4 Max.
- One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes.
- At 689M parameters it scores 83.7% on olmOCR-bench and reaches 4.52 pages per second on one H100 with vLLM.
Key differences
| Qianfan OCR | Surya | |
|---|---|---|
| Licence | Apache-2.0, commercial use allowed | Code Apache-2.0; weights under a modified AI Pubs Open Rail-M, free for research, personal use and companies under $5M funding or revenue |
| Model size | 4B (Qwen3-4B language model) | 689M |
| How you run it | Self-host the weights, or use the Baidu Qianfan platform. Benchmarked on one A100 with vLLM; CPU-only is not a fit | Self-host only. A client runs against a separate vLLM or llama.cpp server, on an NVIDIA GPU, CPU or Apple Silicon |
| Output formats | Markdown, JSON, HTML, LaTeX, bounding boxes | JSON, HTML, LaTeX, bounding boxes, plain text |
| Languages | 192 languages, plus handwriting and scene text | 90+ languages; 74.9% pass rate on a 91-language test, and only 21 of 91 languages reach 90% |
| Speed | 1.024 pages per second on an A100 with W8A8; thinking mode can emit up to 16,384 tokens | 4.52 pages per second at 192 DPI on one H100 with vLLM; 0.198 pages per second on an M4 Max with llama.cpp Q8_0 at 96 DPI |
| Known failure | Slower than MinerU 2.5; PaddleOCR-VL 1.5 scores higher at 94.50 on OmniDocBench v1.5 | Surya 2.5 table output lacks row and column geometry; multilingual tables score 58.1; the README sends handwriting, complex tables and math to the hosted Datalab API |
Sources
What each one is
Qianfan OCR
Baidu's OCR-tuned vision model. Strong on Chinese documents and tables.
Qianfan OCR detailsSurya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya 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": "baidu/qianfan-ocr","fallbacks": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Qianfan OCR alternativesSurya alternativesBest OCR models in 2026