LightOnOCR-3: read the page and its layout

Transcript

The original narration, preserved as published. Relative dates refer to the episode’s coverage window.

A scanned page can hide a whole workflow. LightOn released LightOn O C R three yesterday, October eighth, with downloadable models for reading documents and locating what is on the page.

The family comes in point eight billion, one billion, and four billion parameter sizes. The one billion version keeps the previous architecture. The other two use Qwen three point five vision language models.

An empty prompt gives you a transcription. Use the grounding prompt and you also get labeled boxes, short image descriptions, and chart data as tables. That connects extracted content to its place on the page.

The models are available on Hugging Face under Apache two point zero. LightOn also publishes conversion tools and benchmark scripts. Its reported quality and speed depend on the documents, image resolution, and serving setup.

For a document app, try a messy page you already understand: columns, a chart, small print. Compare the words, reading order, and extracted values against the original. The useful result is a page you can actually work with.