Surya vs InternVL3.5
Pick Surya when you need layout, reading order and bounding boxes for scanned pages, or when you want OCR on a laptop with no big GPU. Pick InternVL3.5 when the job is document question answering or you want one model family you can run at several sizes under Apache-2.0. The main split is output: Surya returns structured blocks with coordinates, while InternVL3.5 returns plain text, Markdown or JSON and no boxes by default.
Suryadatalab/surya | InternVL3.5opengvlab/internvl | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.5 s / page | 3.0 s / page |
| Best for | Layout and line-level OCR | Self-hosted document understanding |
| Inputs | PDF, images | images |
| Languages | 90+ languages | Multilingual |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | No | ✓ Yes |
| Bounding boxes | ✓ Yes | No |
| 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 Surya if
- It returns layout labels, reading order, OCR, HTML tables and KaTeX math with per-block bounding boxes from one model, in 90+ languages.
- It runs on CPU and Apple Silicon through llama.cpp with peak RAM of about 4-6 GB on an M4 Max, and on an NVIDIA GPU through vLLM.
- It has published numbers: 83.7% on olmOCR-bench and 4.52 pages/s at 192 DPI on one H100 with vLLM.
Choose InternVL3.5 if
- It comes in nine sizes from 1B to 241B-A28B, so you can trade accuracy for cost within one family.
- The licence is plain Apache-2.0, with no revenue or funding threshold.
- Models up to 30B fit on a single A100, and the 20B-A4B and 30B-A3B MoE variants activate few parameters per token.
Key differences
| Surya | InternVL3.5 | |
|---|---|---|
| Licence | 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. | Apache-2.0 |
| Size | 689M parameters | 1B to 241B-A28B; the 8B variant is 8.5B |
| How you run it | Lightweight client against a separately running vLLM or llama.cpp server. NVIDIA GPU, or CPU and Apple Silicon via llama.cpp. | Transformers, LMDeploy or vLLM. Up to 30B on one A100, 38B on two A100s, 241B-A28B on eight A100s. |
| Output | JSON, HTML, LaTeX, bounding boxes, plain text | Plain text, Markdown, JSON. No bounding boxes or layout output by default. |
| Benchmarks | olmOCR-bench 83.7% (Surya 2.5), OmniParseBench 85.9, 91-language pass rate 74.9%. Reported by the authors. | Not published as text. OCR scores appear only as chart images on the model card. |
| Tables | Surya 2.5 table output lacks row and column geometry. Multilingual tables score 58.1. | Not published |
| Known limits | Not for handwriting, complex multilingual tables or photos of natural scenes. The README sends hard documents to the hosted Datalab API. | Not suited to jobs that need word coordinates, or CPU-only use of larger sizes. |
Sources
What each one is
Surya
Open-source OCR toolkit with line detection, layout analysis, reading order and tables.
Surya detailsInternVL3.5
Shanghai AI Lab's open vision-language family (1B to 241B, August 2025), strong on OCR and document benchmarks.
InternVL3.5 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": "datalab/surya","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Surya alternativesInternVL3.5 alternativesBest OCR models in 2026