Surya OCR

Surya is an open-source document OCR and analysis toolkit from Vik Paruchuri, the developer behind Marker, built around a 650-million-parameter vision language model. It performs line-level text detection, text recognition across more than 90 languages, layout analysis that classifies tables, images, and headers, reading order detection, and table recognition with cell-level geometry, and it can render recognized math as LaTeX inside HTML output. The recognition model pairs a modern vision-language architecture with a separate EfficientViT-based text line detector, and inference runs through vLLM on GPUs or llama.cpp on CPU and Apple Silicon, turning images and PDFs into JSON with polygon coordinates and confidence scores. The project reports leading accuracy on olmOCR-bench among models of its size. The code is open source, while model weights use a modified Rail license that is free for research, personal projects, and smaller companies, with commercial licensing available beyond that threshold. It suits teams digitizing documents at scale.