Surya

Overview

Surya is a 689M-parameter OCR model from Datalab that performs text recognition, layout analysis, reading order, and table recognition. It supports 90+ languages and is designed for document intelligence workloads across varied document types.

Key Features

• 689M-parameter OCR model with text recognition and line-level text detection.
• Layout analysis for tables, images, headers, and other document elements, with reading order.
• Table output as HTML, including merged cells as colspan and rowspan.
• Multilingual support, scoring 74.9% on an internal benchmark set of 91 languages.
• Throughput of 4.5 pages/s on one H100 at 192 DPI and 1.7x Surya 2 with MTP.

Use Cases

• OCR of images, PDFs, or folders of images and PDFs into JSON with detected text and bounding boxes.
• Layout analysis and reading-order extraction for complex pages such as newspapers, forms, and academic papers.
• Table recognition and HTML table output for documents with merged cells.
• Multilingual document processing across 90+ languages.

Getting Started

• Install with pip install surya-ocr.
• For NVIDIA GPU inference, install Docker and the NVIDIA Container Toolkit; for CPU or Apple Silicon, install the llama-server binary from llama.cpp.
• Run surya_ocr DATA_PATH to write JSON with detected text and bounding boxes; DATA_PATH can be an image, PDF, or folder of images or PDFs.
• Optionally install streamlit and pdftext, then run surya_gui for an interactive app.

Deployment & Requirements

• NVIDIA GPU inference requires Docker plus the NVIDIA Container Toolkit.
• CPU or Apple Silicon inference requires the llama-server binary from llama.cpp.
• Surya auto-spawns the inference server on first use and can also point at an existing server via SURYA_INFERENCE_URL=http://host:port/v1\.

Before You Adopt

• License: Apache-2.0. Review its terms before using, modifying, or distributing the project.
• CPU or Apple Silicon inference involves llama.cpp.
• By default each command starts and stops the VLM server, so repeated commands incur startup and, on GPU, model-load costs.