lightonai/LightOnOCR-2-1B
LightOnOCR-2-1B is LightOn's 1 billion parameter vision-language model designed for efficient, end-to-end Optical Character Recognition (OCR). This model converts documents like PDFs, scans, and images into clean, naturally ordered text without external pipelines. It excels at handling complex layouts including tables, receipts, multi-column documents, and mathematical notation, achieving state-of-the-art performance on OlmOCR-Bench while being significantly faster and more efficient than competing OCR solutions.
Loading preview...
LightOnOCR-2-1B: Efficient End-to-End OCR
LightOnOCR-2-1B is LightOn's flagship 1 billion parameter vision-language model, specifically engineered for robust Optical Character Recognition (OCR). It processes documents such as PDFs, scans, and images directly into structured text, eliminating the need for brittle external OCR pipelines. The model leverages RLVR training for enhanced accuracy and is a refined version of LightOnOCR-2, featuring a larger and higher-quality training corpus with improved coverage for French, arXiv documents, and LaTeX handling.
Key Capabilities & Differentiators
- End-to-End Processing: Fully differentiable, integrating vision and language for seamless text extraction.
- High Speed & Efficiency: Demonstrates significant speed advantages, being 3.3x faster than Chandra OCR and 1.7x faster than OlmOCR. It can process approximately 5.71 pages/second on a single H100 GPU, making it highly cost-effective.
- Versatile Document Handling: Capable of accurately extracting text from diverse and complex layouts, including tables, receipts, forms, multi-column documents, and mathematical notation.
- State-of-the-Art Performance: Achieves leading results on the OlmOCR-Bench benchmark, outperforming larger and slower alternatives.
- Image Detection: Bounding box variants (e.g.,
LightOnOCR-2-1B-bbox) can predict bounding boxes for embedded images within documents.
Good For
- Document Digitization: Converting various document types (scans, PDFs, images) into machine-readable text.
- Information Extraction: Accurately extracting data from structured and semi-structured documents like invoices, receipts, and forms.
- Scientific & Technical Documents: Improved handling of LaTeX and complex mathematical expressions.
- Fine-tuning: The
LightOnOCR-2-1B-basevariant is recommended for domain adaptation (e.g., specific receipt types, scientific articles) and multilingual fine-tuning using LoRA.
For detailed benchmarks and methodology, refer to the official paper.