Neural ExpeditionSubscribe
TEST RECORD · ← all field reports

FIELD REPORTS/FIELD NOTE/ISSUE #75

Turn any document page into clean Markdown, tables included

Document parsing has long meant chaining tools: one model finds the layout, another reads the text, a third rebuilds the tables, and errors compound at every step. This release is the first single model to top the main document parsing leaderboard on its own, ahead of those multi-stage systems. And it does it at a size small enough to run on very modest hardware.

MODELOvisOCR2
PUBLISHEDJuly 16, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYFIELD NOTE

Field notes

01What it does

You give it a picture of a document page, a scan, a PDF page, or a photo, and it writes out the whole page as Markdown in natural reading order. Tables come back as proper HTML tables instead of scrambled text, formulas come back as LaTeX, and figures are marked with their exact position on the page so you can crop and keep them. The output is ready to paste into notes, docs, or a knowledge base without reflowing paragraphs by hand.

The model is a 0.8B build on Qwen3.5 from Alibaba's Ovis team, and it scored 96.58 on OmniDocBench v1.6, the widest document parsing test suite. That made it the first single end-to-end model to lead that board, which pipeline systems had owned until now. For you that means one model call replaces a whole parsing stack, with less to install and fewer places for errors to creep in.

02How to try it

Open the demo Space and feed it your ugliest real page: a dense table, a two-column layout, or a page mixing text with equations. Check whether the table structure survives intact, because that is where OCR tools usually fall apart and where this one is strongest.

Running it locally takes a GPU with vLLM and about thirty lines of Python from the model card. At 0.8B parameters it fits comfortably on modest hardware, so batch-parsing a large document archive is realistic without cloud costs.

03Caveat

The authors are upfront that complex real-world pages can still produce wrong or incomplete output: repeated lines, broken table structures, or shuffled reading order on unusual layouts. For contracts, invoices, or anything where a wrong cell matters, keep a human check on the output.

04What you can do with it

  • Convert paper reports and invoices into editable, searchable Markdown.
  • Pull tables out of PDFs without retyping a single cell.
  • Turn textbook or paper pages with equations into LaTeX you can reuse.
  • Feed clean, structured text into a document Q&A or RAG workflow.
  • Batch-digitize a scan archive into Markdown files with figures preserved.

Try the demo

View model page

Read this issue on neuralexpedition.com →