LightOnOCR-3 — Document Parser Family with Layout Grounding + Chart Data Extraction in One Model, Apache-2.0
Three sizes (0.8B & 4B on the Qwen3.5-VL architecture; 1B on the older LightOnOCR-2 architecture — migrate without changing prompts), all **Apache 2.0**. Two modes: (a) **transcription** — empty prompt, outputs full-page text like the previous generation (generation 2 has been downloaded 4.5 million+ times); (b) **grounding**…
Three sizes (0.8B & 4B on the Qwen3.5-VL architecture; 1B on the older LightOnOCR-2 architecture — migrate without changing prompts), all **Apache 2.0**.
Two modes: (a) **transcription** — empty prompt, outputs full-page text like the previous generation (generation 2 has been downloaded 4.5 million+ times); (b) **grounding** — the `grounding` prompt returns each block with a label + a bounding box normalized to 0–1000: text blocks contain the transcription, image blocks contain a ~10-word description, **and chart blocks contain an HTML table of data points extracted from the graph** (strict rules: printed values are copied verbatim, and values that are not prin…