Ready prompts
AI news

LightOnOCR-3 reads text and page structure in one model

· KGAM.Blog

LightOn released LightOnOCR-3 on October 8, 2026. The family transcribes text, returns document-region coordinates and labels, describes images and extracts numerical data from charts.

A handy generator of ready-to-use prompts
Illustration · KGAM.Blog
Illustration

Models come in 0.8, 1 and 4 billion parameter sizes under Apache 2.0. Code and a demo are available, offering different compute and quality tradeoffs.

Illustration · KGAM.Blog
Illustration

Archives may benefit from a simpler document-processing pipeline. Still compare chart values and complex tables with the original: neat markup does not guarantee correct numbers.