LightOn released LightOnOCR-3 on October 8, adding document layout and visual content extraction to its OCR family. The three sizes are 0.8B, 1B and 4B parameters. Alongside text, the models can produce labeled document regions, descriptions of images and numeric data from charts. LightOn’s launch post
For a document-search system, the useful question is whether this output preserves enough structure to trace an answer back to the page. A chart caption alone may not contain the value a reader needs; a table of extracted values could make it retrievable, provided those values are checked.
Two output modes
The 1B model card documents plain transcription with an empty prompt and a grounding mode that adds labels and bounding boxes. Coordinates are normalized to a 0–1000 page scale. Chart blocks contain extracted data in HTML tables, while image blocks pair regions with descriptions.
The 1B variant retains the LightOnOCR-2 architecture. The 0.8B and 4B variants use Qwen3.5’s vision-language architecture, so deployment choices should follow the selected checkpoint’s instructions rather than assuming all three load identically. The 1B card identifies Apache 2.0 licensing. Model architecture, usage and license
Read the benchmark qualifications
LightOn’s launch comparison notes that the LightOnOCR-2 result excludes headers and footers while the other models include them. Treat that as a limit on direct comparisons between generations. A higher aggregate number alone does not establish how much your own extraction pipeline will improve. Benchmark notes
S5 Labs has not benchmarked these models. Our suggested evaluation is to keep a small, fixed set of documents covering the failures that matter to your application: a multi-column page, a scanned table, a chart with a legend and a page with repeated headers. Compare both the text and its association with the correct region.
Keep evidence attached to extracted data
Store the source page and region alongside each extracted passage. When reviewing a chart, check the axis scale, units and series labels against the image before accepting the output table. An accurately transcribed number attached to the wrong series is still a factual error.
For retrieval, compare answer quality before replacing the existing parser. Count missing passages, incorrect reading order and unsupported answers separately. This gives you a way to determine whether richer extraction improves the result or simply adds more material to index. The model’s output should remain traceable to the document a reader can inspect.