AgentPlays
LightOnOCR-3: Apache-2.0 OCR models that also return layout boxes and chart tables
LightOn released LightOnOCR-3 in 0.8B, 1B and 4B sizes under Apache 2.0. With an empty prompt it transcribes a page; with the "grounding" prompt it also returns labeled bounding boxes for each region, short image descriptions and chart data as HTML tables. Speed figures are LightOn's own: it says the 0.8B peaks at 4.78 pages/s on one H100 and the 4B serves 21% more pages per second than Chandra-OCR-2. LightOn doesn't publish a cost figure. A free demo runs on Hugging Face.
How you’d use it: Turn scanned PDFs into text plus layout boxes and chart tables on your own GPU.