The Pulse
LightOn Releases OCR-3 Models That Read Layouts, Charts and Images
LightOn released three open OCR-3 models that transcribe documents and can also locate page elements, describe images and extract chart data. The models use the Apache 2.0 license, and LightOn's benchmarks show strong results alongside a ga
AI.info Team · Written with AI ·
LightOnOCR-3 promises to replace parts of the document-processing stacks that combine separate tools for text, page layout and charts. But LightOn’s own benchmark results show a more qualified picture: its largest model trails the top scorer on one OCR test, even as the family posts strong results on others.
LightOn announced the models on October 8, 2026, offering three sizes: 0.8 billion, 1 billion and 4 billion parameters. All are released under the Apache 2.0 license for research and commercial use. The company says the new models add layout and visual analysis to text transcription, targeting the extra processing often needed to make complex documents searchable and usable by downstream applications.
One prompt adds coordinates, image descriptions and chart data
With an empty text prompt, LightOnOCR-3 transcribes a page using the interface of earlier releases. A prompt containing only “grounding” requests a richer output: labeled bounding boxes for page elements, short descriptions for images, and chart data represented as HTML tables. Coordinates use a 0–1,000 scale, according to LightOn’s technical post.
The distinction matters for documents where meaning depends on more than extracted words. Keeping paragraph, table and image locations may help document-search systems preserve context, while turning a chart into a table can make its values available to software that would otherwise receive only a picture. LightOn pitches the combined output as a way to reduce reliance on separate layout-detection and chart-processing steps.
Strong scores, with a clear gap on olmOCR-Bench
On olmOCR-Bench, LightOn reports an overall score of 86.3 for its 4B model, 1.3 points behind Infinity Parser Pro, which lists 35.1 billion parameters. The LightOn model scores 0.5 points above Chandra 2, another 4B model, in the company’s comparison. Its smaller variants score 85.5 for the 0.8B model and 84.5 for the 1B model.
Results vary by document type. LightOn says the 4B version leads on ArXiv pages, while the 1B model leads on multi-column pages and the 0.8B version on long, tiny text. Infinity Parser Pro remains ahead on old scans and headers and footers. On ParseBench, LightOn’s 4B and 0.8B models score 75.1 and 74.6 on the five-category overall score, respectively, narrowly ahead of Infinity Parser Pro’s 74.3; on the French-document FRBench-pdf2md test, the 4B model scores 74.1.
LightOn flags benchmark scoring and speed trade-offs
The scores come with a caveat: LightOn says the benchmarks use edit distance, which can penalize equivalent formatting choices, such as representing a footnote in HTML rather than LaTeX. The company says it applies normalization functions to competing models’ outputs and publishes those functions and reproduction scripts. That qualification matters when reading small score differences as evidence of better extraction.
LightOn also reports that the 0.8B and 4B models produce 9% to 14% fewer output tokens on average than Chandra-OCR-2 and Infinity-Parser2-Pro across the same 512 olmOCR-bench pages in grounding mode. Its speed tests used one H100 per model and vLLM 0.30.0; the company says rendering pages at lower resolution boosts peak throughput, while the 1B version has the lowest single-page latency at 2.7 seconds.
LightOn’s case for OCR-3 rests on putting transcription and visual structure into one model call, not on beating every competitor in every category. The three Apache 2.0 model checkpoints are available on Hugging Face, and the company has also released scripts for converting and visualizing the grounded output.