LightOnOCR is LightOn's family of small, open OCR models that turn pictures of document pages into text, with version 3 also mapping where each part of the page sits.
In plain words
OCR (optical character recognition) reads the words in a scanned or photographed page. Given no prompt, a LightOnOCR model writes out the whole page as text. LightOnOCR-3, released in October 2026, also boxes and labels every block on the page, describes each picture briefly and pulls the numbers out of charts, so one model can replace a chain of document tools. It comes in 0.8, 1 and 4 billion parameter sizes, all open under the Apache 2.0 licence for research and commercial use.
A quick example
Imagine a shop scanning a pile of paper invoices. The model returns each invoice's text, marks where the table and the totals sit on the page, and turns a bar chart in the footer into rows of numbers the shop can search.
Where you'll meet it
- Turning scanned documents into searchable text.
- Feeding page text, layout and chart data to search and question-answering tools.
Related
Sources
- Docslightonai/LightOnOCR-3-1B, LightOn (Hugging Face) · read 8 Oct 2026
- DocsLightOnOCR-3: High-Performance OCR and Layout Extraction in One Model, LightOn (Hugging Face blog) · read 8 Oct 2026
Checked by WhatIsAI against the sources above · 8 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.