LightOn AI Releases LightOnOCR-3-4B Open-Weights Model for Document OCR
French AI startup LightOn has released LightOnOCR-3-4B on Hugging Face, a 3.4 billion parameter open-weights model specialized for optical character recognition (OCR) and document understanding. Compact, specialized open-weights models like LightOnOCR-3-4B allow developers to parse dense printed documents and scientific PDFs locally without relying on costly proprietary APIs. This helps organizations process sensitive visual documentation securely within their own infrastructure. LightOnOCR architecture combines a Pixtral-based Vision Transformer encoder with a lightweight Qwen-based text decoder distilled from larger vision-language models. The model is optimized for extracting dense typography, complex table structures, and LaTeX mathematical formulas.
## BACKGROUND
Optical Character Recognition (OCR) is the process of extracting editable, machine-readable text from scanned documents or images. Modern document understanding uses Vision-Language Models (VLMs) that pair image encoders with text decoders to preserve formatting, tables, and mathematical markup during extraction.