Extract document text, layout and chart data with LightOnOCR-3
Added to DistroCool on 2026-10-20 · Discovery date, not a release announcement
LightOn's October 8, 2026 release combines page transcription with labeled bounding boxes, image descriptions and chart-to-table extraction. The Apache 2.0 family includes 0.8B, 1B and 4B variants.
Why it matters
Document-processing builders can preserve where information appeared on a page instead of receiving only a flat text transcript. This is an extraction model, not a guarantee that every number or reading order is correct.
Try this next.
Use the official demo with a public sample page, or follow the model card's Transformers setup. Compare an empty prompt for transcription with the grounding prompt for layout and chart data. Check every extracted number; self-hosting needs model downloads and suitable compute. DistroCool has not independently tested this example.
Source: LightOn team release and usage guide ↗