Sarashina2.2-OCR — OCR AI Model
Sarashina2.2-OCR reads Japanese and English documents back as markdown, with tables as HTML and formulas as LaTeX. It handles vertical Japanese text, which the other OCR models do not. Figures and charts come back marked with a box on the image output. Seven to twenty seconds depending on page density. Japanese and English only — for Korean use PaddleOCR, for other languages GLM-OCR.
What it does
Sarashina2.2-OCR reads a Japanese or English document back as markdown, keeping the headings, tables and formulas rather than flattening the page into a wall of text.
It is the only OCR model here that reads vertical Japanese text, and it keeps the right-to-left column order a Japanese newsletter or government page is written in. Tables come back as HTML, formulas as LaTeX, and any figures or charts are marked with a box on the image output. Japanese and English only: for Korean, use PaddleOCR; for other languages, GLM-OCR.
When to use it
- A Japanese document set vertically, read in the order it was written
- Municipal newsletters, policy slides and government forms
- Tables kept as tables instead of collapsing into loose lines
- Mathematical formulas kept as LaTeX rather than mangled text
- Finding where the figures and charts sit on a page
Details
- CategoryOCR
- Year2026
- LicenseMIT
Compliance & Provenance
- ProviderOpen-source (SB Intuitions) · Specialized
- EU AI Act RiskMinimal Risk
- Art. 50 TransparencyNot applicable
Inputs & Outputs
- ImageInputimage
Source document image containing text to recognize
- ImageOutputimage
Detected figures and charts marked as an image
- TextOutputstring
Detected texts as markdown
Tags
Alternatives in Text Recognition(OCR)
- DeepSeekOCR
Multi-language + handwriting OCR. 30-60s.
- GLM-OCR
Document OCR with layout/table/formula. 10-40s.
- PaddleOCR
Printed-text OCR for Korean, Japanese, Chinese and English. Fast.