Typhoon OCR vs RolmOCR
Pick Typhoon OCR for Thai and English documents such as government forms, financial reports and books, where you need tables as HTML and equations as LaTeX. Pick RolmOCR for plain text transcription of rotated or skewed scans in English-style pipelines on one GPU. The main split is output: Typhoon OCR returns structured Markdown, HTML and LaTeX, while RolmOCR returns plain text only.
Typhoon OCRtyphoon/typhoon-ocr | RolmOCRreducto/rolmocr | |
|---|---|---|
| Type | Open source | Open source |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 1.8 s / page | 3.0 s / page |
| Best for | Thai documents | Fast open PDF to text |
| Inputs | PDF, images | PDF, images |
| Languages | Thai, English | Not stated by the authors |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | ✓ Yes |
| Handwriting | No | ✓ Yes |
| Math | No | No |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | 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 Typhoon OCR if
- It is tuned for Thai: V2 2B scores 81.3 on MDPBench Thai pages versus 61.3 for its base model, and V1.5 beat GPT-5 and Gemini 2.5 Pro on Thai government forms (BLEU 0.870).
- It outputs Markdown, HTML and LaTeX, with tables as HTML including merged cells, equations as LaTeX and checkboxes marked.
- It is small at 2B parameters yet scores 84.7 on olmOCR-bench, and it can be self-hosted or called through the Typhoon API as typhoon-ocr-v2.
Choose RolmOCR if
- It is built for rotated and skewed scans, since about 15% of its training data was rotated to handle off-angle pages.
- It is a straightforward self-hosted plain text transcriber under Apache-2.0, served for example with vLLM on one GPU.
- Reducto reports faster inference and lower memory use than the original olmOCR, and Needl.ai counted 7 hallucination issues versus 11 for olmocr-2-7b.
Key differences
| Typhoon OCR | RolmOCR | |
|---|---|---|
| Licence | Apache-2.0, subject to OpenTyphoon Terms and Conditions | Apache-2.0, usable commercially |
| How you run it | Self-host the weights or use the Typhoon API (typhoon-ocr-v2) | Self-host only, for example with vLLM |
| Size | 2B (V2 and V1.5); 0.8B V2 variant | 8B (Qwen2.5-VL-7B base) |
| Output formats | Markdown, HTML and LaTeX; tables as HTML with merged cells | Plain text only; no layout bounding boxes |
| Benchmarks | olmOCR-bench 84.7, OmniDocBench v1.6 90.97, ThaiOCRBench 59.9 (V2 2B, reported by the authors) | Needl.ai PDF extraction overall 0.663, versus 0.770 for olmocr-2-7b (independent) |
| Languages | Thai and English only evaluated | Not published |
| Known failure | V2 can loop until the token limit, mostly on dense multi-column pages; key-information extraction is weakest at 21.0 on ThaiOCRBench | Can hallucinate or drop content, such as a table subtitle that olmOCR extracted; quantized versions not evaluated |
Sources
- Typhoon OCR V2 2B model card
- Typhoon OCR technical report
- Hugging Face model card
- Reducto announcement
- Needl.ai benchmark
What each one is
Typhoon OCR
Open vision model tuned for Thai and English documents, with Markdown output.
Typhoon OCR 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": "typhoon/typhoon-ocr","fallbacks": ["reducto/rolmocr"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Typhoon OCR alternativesRolmOCR alternativesBest OCR models in 2026