GOT-OCR 2.0 vs InternVL3.5
Pick GOT-OCR 2.0 when you need a small model that reads formulas, sheet music, chemical structures or a chosen region of a page, and you can filter blank pages. Pick InternVL3.5 when you want document question answering with JSON or Markdown output, an Apache-2.0 licence and a choice of nine sizes. The main split is a narrow OCR specialist at 580M parameters against a general vision language model that scales from 1B to 241B-A28B.
GOT-OCR 2.0stepfun/got-ocr-2 | InternVL3.5opengvlab/internvl | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.5 s / page | 3.0 s / page |
| Best for | Formatted text, formulas and charts | Self-hosted document understanding |
| Inputs | images | images |
| Languages | Chinese, English | Multilingual |
| Max pages per request | 100 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | No |
| Math | ✓ Yes | ✓ Yes |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | ✓ 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 GOT-OCR 2.0 if
- It outputs LaTeX, TikZ, SMILES and kern, so it covers formulas, chemical structures and sheet music.
- Region OCR takes box coordinates or color, so you can read one part of a page.
- At 580M parameters it is a single small model, and it is merged into Hugging Face Transformers with batched inference.
Choose InternVL3.5 if
- Nine sizes from 1B to 241B let you trade accuracy for cost, and models up to 30B fit on a single A100.
- It is a good fit for document question answering and returns plain text, Markdown or JSON.
- The licence is Apache-2.0 and it is built on Apache-2.0 Qwen3, with no non-commercial term attached.
Key differences
| GOT-OCR 2.0 | InternVL3.5 | |
|---|---|---|
| Licence | Code Apache-2.0; training data CC BY-NC 4.0 | Apache-2.0 |
| Model size | 580M parameters, one model | 1B to 241B-A28B in nine sizes; the 8B variant is 8.5B |
| Output formats | Plain text, Markdown, LaTeX, TikZ, SMILES, kern | Plain text, Markdown, JSON |
| Hardware | CUDA 11.8 and PyTorch 2.0.1 per the README; CPU users are pointed to a llama.cpp port | Up to 30B on a single A100; 38B needs two A100s; 241B-A28B needs eight A100s |
| Layout and coordinates | Region OCR guided by box coordinates or color | No bounding boxes or layout output by default |
| Benchmark scores | Not published in the README | Not published as text; OCR scores appear only as chart images |
| Known failure | Can loop characters midway, return wrong text on images with no text, and does worse in format mode than in plain OCR mode | Large sizes need multi-GPU setups, and it is not suited to jobs that need word coordinates |
Sources
What each one is
GOT-OCR 2.0
Unified OCR model that reads plain text, tables, formulas, sheet music and charts.
GOT-OCR 2.0 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": "stepfun/got-ocr-2","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
GOT-OCR 2.0 alternativesInternVL3.5 alternativesBest OCR models in 2026