Table formats: how OCR engines write a table

How 14 OCR engines write a table into their output (HTML, Markdown or another format), whether merged cells survive, and whether you can choose. From each engine's official docs, model card or code.

Key takeaways

• Landing AI made HTML its default table format on Jun 26 2026; HunyuanOCR-1.5, OvisOCR2 and dots.ocr also write tables as HTML inside their Markdown.
• Mistral OCR 4.1 still defaults to Markdown tables, which can't hold a merged cell; its table_format option returns HTML instead.
• Qianfan-OCR's card names HTML only for its table-recognition task: how a full page's tables come back needs a test run. PaddleOCR-VL-1.6's own pipeline code writes them as HTML.

The scoreboard

How it was measured

Each row comes from the engine's official API docs, model card or README, and for PaddleOCR-VL-1.6 from its pipeline code (PaddleX 3.7.2), read on Oct 3 2026; the Source column dates it. Not stated means the official page doesn't say, and nothing is guessed. A Markdown pipe table can't span a cell across columns or rows; HTML can, with colspan and rowspan. In our own test on Oct 2 2026, a Closing Disclosure cost table read by Mistral OCR 4.1 as Markdown put $1,711.00 under At Closing only; as HTML, it spanned At Closing and Before Closing, as on the page.

Log