GOT-OCR 2.0 vs RolmOCR
Pick GOT-OCR 2.0 for formulas, sheet music, chemical structures and region-specific OCR on English and Chinese documents, where you need LaTeX, Markdown or other structured output. Pick RolmOCR for plain text transcription of scanned pages, especially rotated or skewed ones, under an Apache-2.0 licence for the weights. The main split is output type: GOT-OCR 2.0 reads many content types, while RolmOCR only returns plain text.
GOT-OCR 2.0stepfun/got-ocr-2 | RolmOCRreducto/rolmocr | |
|---|---|---|
| Type | Open source | Open source |
| 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 | Fast open PDF to text |
| Inputs | images | PDF, images |
| Languages | Chinese, English | Not stated by the authors |
| Max pages per request | 100 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | No |
| 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
- You need formulas, sheet music or chemical structures, with output as LaTeX, TikZ, SMILES or kern.
- You want to OCR one region of a page, guided by box coordinates or color.
- You have limited GPU memory: it has 580M parameters, and a llama.cpp port exists for CPU users.
Choose RolmOCR if
- You are transcribing scanned pages to plain text, including rotated or skewed ones, since about 15% of its training data was rotated.
- You need Apache-2.0 weights that are usable commercially, with no non-commercial data licence attached.
- You have an independent benchmark to check against: Needl.ai scored it 0.663 overall and counted 7 hallucination issues versus 11 for olmocr-2-7b.
Key differences
| GOT-OCR 2.0 | RolmOCR | |
|---|---|---|
| Model size and base | 580M parameters, unified end-to-end OCR model from StepFun and UCAS | 8B (Qwen2.5-VL-7B base), fine-tuned by Reducto on olmOCR-mix-0225 |
| Licence | Code Apache-2.0; training data CC BY-NC 4.0 | Apache-2.0 |
| How you run it | Self-hosted. Needs CUDA 11.8 and PyTorch 2.0.1 per the README; available in Hugging Face Transformers; llama.cpp port for CPU | Self-hosted, for example with vLLM; suited to one GPU |
| Output formats | Plain text, Markdown, LaTeX, TikZ, SMILES, kern | Plain text only |
| Layout and regions | Region OCR by box coordinates or color; multi-page OCR | Cannot output layout bounding boxes |
| Benchmark | Not published (README lists no benchmark scores) | 0.663 overall on Needl.ai (independent), versus 0.770 for olmocr-2-7b |
| Known failure | Output can loop characters midway (issue #226); returns wrong text on blank images (issue #235); format mode is much worse than plain mode (issue #317) | Can hallucinate or drop content; omitted a table subtitle that olmOCR extracted; quantized versions not evaluated |
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 detailsRolmOCR
Reducto's open PDF-to-text model, built on Qwen2.5-VL-7B with the olmOCR recipe. Needs no PDF metadata and handles rotated pages.
RolmOCR 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": ["reducto/rolmocr"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
GOT-OCR 2.0 alternativesRolmOCR alternativesBest OCR models in 2026