OCRRouter

Surya vs Florence-2

Pick Surya for full-page document work: it returns layout labels, reading order, HTML tables and math with bounding boxes in 90+ languages, and it scores 83.7% on olmOCR-bench. Pick Florence-2 for short text, scene text and labels in English where you want a small MIT-licensed model that also does captioning and detection. Florence-2 returns only text and boxes, so it cannot convert a page to structured output.

Suryadatalab/suryaFlorence-2microsoft/florence-2
TypeOpen sourceOpen source
Price2 credits / page1 credit / page
Per 1,000 pages$2.00$1.00
Typical speed1.5 s / page0.8 s / page
Best forLayout and line-level OCRLightweight OCR with boxes
InputsPDF, imagesimages
Languages90+ languagesEnglish
Max pages per request500500
Tables✓ YesNo
HandwritingNoNo
MathNoNo
Bounding boxes✓ Yes✓ Yes
ConfidenceNoNo
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 Surya if

  • You need structured output from a page: layout labels, reading order, HTML tables and KaTeX math, with per-block bounding boxes.
  • You work in many languages: Surya covers 90+ languages, while Florence-2 gets Chinese output mostly wrong because its vocabulary lacks CJK characters.
  • You want to run locally on a CPU or Apple Silicon laptop through llama.cpp, with peak memory of about 4.3-5.9 GB on an M4 Max.

Choose Florence-2 if

  • You want a very small model: 0.23B base or 0.77B large, under an MIT licence with no revenue or funding cap.
  • You need quadrilateral boxes per text line on scene text or short labels, using the OCR_WITH_REGION task.
  • Your pipeline also needs captioning, detection or grounding from the same model, and it needs only plain text and boxes.

Key differences

SuryaFlorence-2
LicenceCode Apache-2.0; weights under a modified AI Pubs Open Rail-M, free for research, personal use and companies under $5M funding or revenueMIT
Model size689M parameters0.23B (base), 0.77B (large)
How you run itLightweight client against a vLLM or llama.cpp inference server, which you must run separately; NVIDIA GPU via vLLM, or CPU and Apple Silicon via llama.cppSelf-hosted via Transformers
Output formatsJSON, HTML, LaTeX, bounding boxes, plain textPlain text and bounding boxes only; no Markdown, tables or LaTeX
Benchmark83.7% on olmOCR-bench (Surya 2.5), as reported by the authorsNo OCR benchmark published; reports TextCaps CIDEr 72.8 and COCO detection mAP 37.5 (zero-shot, large)
Languages90+ languages; 74.9% pass rate on a 91-language test, with 21 of 91 languages reaching 90%Chinese output mostly wrong; spelling errors and missing spaces seen on French text
Handwriting and known failuresREADME sends handwriting, complex tables and math to the hosted Datalab API; Surya 2.5 table output lacks row and column geometry; not for photos or natural scenesSpelling errors and missing spaces on cursive handwriting; the continued-pretrained checkpoint used only 0.1B samples and might not be well trained

Sources

What each one is

Surya

Open-source OCR toolkit with line detection, layout analysis, reading order and tables.

Surya details

Florence-2

Small Microsoft vision model (0.23B and 0.77B) with OCR and OCR-with-regions tasks that return text with boxes.

Florence-2 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/surya","fallbacks": ["microsoft/florence-2"],"file_url": "https://example.com/document.pdf"}

Questions

More comparisons

Surya alternativesFlorence-2 alternativesBest OCR models in 2026