Snip Snipping Tool Chrome Extension OCR API Files API Private Cloud OCR Secure Conversion Service
Make Documents Accessible Process Chemical Documents Collaborate on Documents OCR API for Developers Train Language Models Support Academic Research Artificial Intelligence Fintech Edtech Pharma & Chemical Universities & Schools
Handwriting Recognition Digital Ink On-prem PDF Cloud Mathpix Markdown All Supported Languages Image Conversion PDF Conversion Markdown Conversion Table OCR Mathpix CLI PDF Search PDF Reader PDF Data Extraction Chrome Extension View Conversion Gallery
Snip APIs Private Cloud OCR SCS
Mobile Desktop Web Chrome Extension
Mathpix Snip Apps Mathpix OCR API Mathpix Markdown Python SDK
Blog
About Careers Contact
Contact Get Started
← Back to Blog

SuperNet-202: Better OCR for dense pages, vertical text and forms

2026-11-03 · api, updates
Today we’re releasing SuperNet-202, an update to our OCR model for v3/text, v3/pdf and the Files API. It focuses on recognition accuracy, with a few new symbols and a set of smaller fixes. The improvements apply automatically, with no changes to your requests.

Better recognition

Dense pages such as newspapers are recognized more completely, with less missing text. Vertical Chinese and Japanese text is recognized more accurately on complex vertical layouts. Portuguese ordinal indicators such as º and ª are now recognized on low-resolution and complex pages as well.
Checkboxes and radio buttons are recognized more accurately. Longer lines of text are now recognized as text instead of being returned as images; very long lines are still returned as images.

New symbols

Keyboard shortcuts keep their key symbols, such as ⌘+⇧+3, ⌘+⇧+4 and ⌘+⇧+5.
The shortcuts ⌘+⇧+3 / ⌘+⇧+4 / ⌘+⇧+5 in the OCR output, above the same shortcuts in the source image
The OCR output (top) keeps the key symbols of the source image (bottom).
Up, down and equal indicators in tables, △, ▽ and ◇, are recognized as symbols, including colored and filled versions.
Table cells with green upward and red downward triangle indicators beside their values
OCR output of the same table cells, with the indicators recognized as △ and ▽
Colored indicators in the source table (top) are recognized as △ and ▽ in the OCR output (bottom).

Other fixes

  • Equation numbers are no longer lost, and adding them no longer produces invalid LaTeX.
  • Short, single-line figure captions no longer run into nearby text.
  • Lines in table cells on curved pages stay in reading order.
  • Diagonally split table cells keep both of their labels.
  • In a table of contents, a page number follows its title when the title wraps onto a second line.
  • Paragraphs in table footnotes stay separate.
  • Region coordinates no longer extend past the page edge, and regions outside the page are no longer returned as empty images.
  • Fixed a rare case where the same region was returned many times.

Change to line data

Page columns are no longer returned as separate regions. Line data no longer includes entries of type column, and the column field of every line in lines.json and lines.mmd.json is now 0. Lines are still listed in reading order, each with its region, so columns can be derived from line positions. Table columns are unchanged.
See the changelog for release details. Please reach out to support@mathpix.com if you have any questions.