How the PDF pipeline works
OCRZero runs four steps for every PDF.
- Render. Each page is drawn to an image at 150 dpi; wide pages are scaled down to 2,200 pixels.
- Parse. Xiaomi-OCR-0 turns every page image into Markdown in reading order, ignoring headers and footers.
- Clean up. The official postprocess converts OTSL tables to HTML, normalises math delimiters and trims runaway repetition.
- Join. Pages are joined in order, each marked with an HTML comment such as page 3.
Tables, formulas and layout
The model writes tables in OTSL, a compact table markup with merged cells. OCRZero converts them to HTML tables with rowspan and colspan, which most Markdown renderers and RAG pipelines accept.
Formulas are written in LaTeX, inline as $ … $ and display as $$ … $$. Figures are not described; Xiaomi-OCR-0 extracts the text content of the page.
Speed and cost
In our tests on a text-heavy research paper, a page took about 5 to 15 seconds on a single stream, depending on how much text it held. On OCRZero every page costs 1 credit: $5 per 1,000 pages, with 20 free pages for new eligible accounts.