LightOn has introduced LightOnOCR-3, a family of three downloadable document-reading models designed to combine optical character recognition with page-layout understanding. Its October 8 announcement identifies 0.8-billion-, 1-billion- and 4-billion-parameter variants and says they are released under Apache 2.0. The update broadens an OCR workflow beyond transcription: the models can also identify document regions with labeled coordinates, describe images and extract numerical content from charts or figures.
The interface makes the distinction explicit. An empty prompt produces traditional text transcription, preserving the earlier LightOnOCR behavior. A 'grounding' prompt requests additional structure: boxes for paragraphs and titles, image descriptions and HTML tables representing chart data. LightOn says coordinates use a normalized 0–1000 scale. This offers downstream document-search and retrieval pipelines an alternative to chaining independent OCR, layout detection and chart-understanding systems, although production suitability still depends on workload-specific validation.
The sizes are not interchangeable from an implementation standpoint. According to the publisher, LightOnOCR-3-1B retains the preceding architecture, while the 0.8B and 4B variants use a Qwen3.5 vision-language architecture. Separate repositories are linked for all three checkpoints. LightOn also provides utilities for converting the output format, visualizing regions and reproducing its evaluation procedures. Those artifacts are more useful evidence than a headline accuracy score alone, because teams can inspect exactly which layout representations and scripts are supplied.
LightOn reports improvements on several OCR and document parsing benchmarks and discusses speed and token-format tradeoffs. Open Model Weights has not independently run those tests or established a universal ranking. The release matters for self-hosted document AI because both the files and the publisher's declared Apache-2.0 terms can be inspected. Before deployment, exact checkpoint licenses, source revisions, runtime compatibility and memory requirements still need artifact-level verification rather than assumptions transferred across the family.
What this changes for the evidence layer
Three separately downloadable checkpoint variants and their exact license files should be assessed as distinct model artifacts, not conflated into a single record with inherited runtime claims. LightOn describes Apache-2.0 licensing in its official release; verify license files and exact repository revisions before classifying individual registry fields. All OCR accuracy, throughput and formatting comparisons remain developer-reported rather than OMW-measured.
LightOn official Hugging Face release article
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗