Unlimited OCR vs Surya
Pick Unlimited OCR for long reports and books that you want parsed in one request on an NVIDIA GPU, with an MIT licence that has no revenue cap. Pick Surya for layout detection, reading order and local OCR on a laptop or CPU, with the licence caveat for companies above $5M in funding or revenue. The main split is long-document parsing on a GPU versus a small model that runs almost anywhere.
Unlimited OCRbaidu/unlimited-ocr | Suryadatalab/surya | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.2 s / page | 1.5 s / page |
| Best for | Books and very long documents | Layout and line-level OCR |
| Inputs | PDF, images | PDF, images |
| Languages | 30+ languages | 90+ languages |
| Max pages per request | 1,000 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | ✓ 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 Unlimited OCR if
- It parses 40+ pages in a single forward pass with roughly constant memory, so a long report or book goes in as one request.
- The MIT licence allows commercial use with no copyleft and no revenue or funding limit.
- Baidu Cloud offers a hosted endpoint, so you can skip self-hosting if you want.
Choose Surya if
- It runs on CPU and Apple Silicon through llama.cpp, with peak RAM of about 4 to 6 GB on an M4 Max, while Unlimited OCR needs an NVIDIA GPU.
- One model returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes, across 90+ languages.
- At 689M parameters it scores 83.7% on olmOCR-bench and runs at 4.52 pages/s at 192 DPI on one H100 with vLLM.
Key differences
| Unlimited OCR | Surya | |
|---|---|---|
| Licence | MIT, free for commercial use | 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 |
| How you run it | Self-host with vLLM, SGLang or transformers, or use the hosted endpoint in Baidu Cloud | Self-host only. A lightweight client talks to a separately running vLLM or llama.cpp server |
| Hardware | NVIDIA GPU with CUDA 12.9 or 13.0. Version 2 under SGLang runs out of memory on 12 GB GPUs. No CPU-only use | NVIDIA GPU via vLLM, or CPU and Apple Silicon via llama.cpp |
| Model size | 3B total (Unlimited OCR 2 is 3.34B) | 689M |
| Output formats | Markdown, plain text, bounding boxes | JSON, HTML, LaTeX, bounding boxes, plain text |
| Benchmarks | ParseBench mean 46.17 (independent). Unlimited OCR 2: MPDocBench 84.6 and MarathonDocBench 77.1 (authors) | Surya 2.5: olmOCR-bench 83.7%, OmniParseBench 85.9, 91-language pass rate 74.9% (authors) |
| Known failures | Text repetition and looping (issue #55), hallucinated text (issue #58), bbox coordinate loops on GB10 GPUs with SGLang (issue #114). Bold and italic formatting is mostly lost. Context capped at 32,768 tokens | Surya 2.5 tables lack row and column geometry. Multilingual tables score 58.1, and only 21 of 91 languages reach a 90% pass rate. Handwriting, complex tables and math are better sent to the hosted Datalab API |
Sources
- Unlimited OCR Hugging Face model card
- Unlimited OCR GitHub repo
- MarkTechPost on Unlimited OCR 2
- Surya GitHub README
- Datalab pricing and license FAQ
What each one is
Unlimited OCR
Open-source model from Baidu that reads hundreds of pages in one pass with flat memory use.
Unlimited 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/unlimited-ocr","fallbacks": ["datalab/surya"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Unlimited OCR alternativesSurya alternativesBest OCR models in 2026