LightOnOCR-3 adds grounded document extraction across three open models

LightOn’s Apache 2.0 OCR family combines Markdown transcription with layout grounding, while its own benchmark analysis highlights normalization and evaluation caveats.

2 min read

A person feeds a document into an office printer and scanner, illustrating OCR and document digitization.

LightOn has released LightOnOCR-3, a three-model family that combines page transcription with layout-aware extraction. The Apache 2.0 release includes 0.8B-, 1B- and 4B-class variants on Hugging Face. The 1B model retains the earlier LightOnOCR architecture, while the 0.8B and 4B versions use Qwen3.5 vision-language backbones.

The models accept either a standard transcription prompt or a grounding prompt. In grounding mode, output can include labeled bounding boxes, image descriptions and chart-to-table conversion alongside Markdown text. That makes the release relevant beyond plain OCR: it targets document pipelines that need both readable text and coordinates for downstream review or extraction.

LightOn reports an 86.3 score for the 4B model on olmOCR-Bench, below Infinity Parser Pro’s 87.6 but above Chandra2’s 85.8. On ParseBench, the company reports 75.1 for the 4B model and 74.6 for the 0.8B model, compared with 74.3 for Infinity Parser Pro. The comparison is not a clean universal ranking. LightOn’s own analysis shows that normalization and post-processing choices can materially change scores, and it separately notes weaker visual-grounding results than Infinity Parser Pro.

The publisher also claims the 4B model processes a single page 19% faster than Chandra and achieves 21% higher page throughput at the tested settings. Those are vendor-run results tied to particular resolutions and serving configurations, so independent replication will matter.

Weights, code and an interactive demo are available now. The release is unusually useful for evaluation because LightOn publishes task-level results, training details and explicit caveats rather than only a headline aggregate.

Sources: LightOnOCR-3 release and evaluation · 0.8B model · 1B model · 4B model

benchmarksOCRLightOnmultimodal AIopen models