lightonai/LightOnOCR-3-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 6, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

LightOnOCR-3-4B is a 4.5 billion parameter vision-language model developed by LightOn, designed for end-to-end Optical Character Recognition (OCR) and document understanding. Built on the Qwen3.5 architecture, it excels at transcribing text, identifying bounding boxes for visual elements, and extracting structured data from images and charts. This model offers a streamlined alternative to complex document processing pipelines, making it ideal for versatile OCR tasks including tables, receipts, forms, and multi-column layouts.

Loading preview...

LightOnOCR-3-4B Overview

LightOnOCR-3-4B is the largest and most accurate model in LightOn's new generation of lightweight, end-to-end OCR models. This 4.5 billion parameter model, based on the Qwen3.5 vision-language architecture, significantly improves transcription quality and speed compared to previous generations. It integrates layout and visual understanding, providing a comprehensive solution for document processing.

Key Capabilities

  • End-to-End OCR: Processes entire documents to extract text and visual information, simplifying deployment and maintenance by replacing complex pipelines.
  • Transcription Mode: Outputs full page text in markdown format when called with an empty prompt, offering a direct upgrade path from LightOnOCR-2.
  • Grounding Mode: When prompted with grounding, it provides transcribed text along with labels and bounding box coordinates for every visual element (text, images, charts).
  • Visual Understanding: Generates short descriptions for images and extracts numerical data from charts, presenting it as HTML tables.
  • Versatile Document Handling: Capable of processing diverse document types including tables, receipts, forms, charts, multi-column layouts, and mathematical notation.

Good for

  • Applications requiring highly accurate and efficient OCR.
  • Automating document understanding workflows where both text and visual context are crucial.
  • Extracting structured data from various document formats, such as financial reports or scientific papers.
  • Developers seeking a single model solution for OCR and document layout analysis, reducing pipeline complexity.

Finetunes
1 models