Surya
github.com
- Free plan
- macOS
- Linux
- API
- Self-hosted
What Surya is
Surya is a free, open-source document OCR model with 650 million parameters. It handles text recognition, page layout, reading order and table recognition, and supports 91 languages. Surya 2 returns page content as structured blocks that can include HTML, polygons, bounding boxes, confidence and reading order. Table recognition can return row and column geometry; full mode can produce HTML that retains spanning cells and header rows. Equations detected in page OCR are represented as KaTeX-compatible LaTeX inside math tags. The project offers a Python package, command-line tools and an interactive Streamlit app for images and PDFs. Layout, OCR and table functions need a running inference backend: vLLM on NVIDIA GPUs or llama.cpp on CPU and Apple Silicon, or a connection to an existing server. Surya is specialized for documents rather than photos or natural scenes. The code is Apache 2.0, while model weights use a modified AI Pubs Open Rail-M license with commercial-use conditions described in the README. Datalab also offers managed access, a public playground and $5 in free credits.
Who it suits
Surya suits developers and teams building document OCR workflows who can run or connect an inference backend. Its document focus and separate model-weight licensing terms are relevant considerations for prospective users.
Good to know, and what to watch
- Supports OCR in 91 languages.
- Recognizes text, layout, reading order and tables.
- Provides Python, command-line and Streamlit interfaces.
- Returns structured page content and table geometry.
- Free and open source.
- Designed for documents, not photos or natural scenes.
- Layout, OCR and table recognition need an inference backend.
- Model weights have separate commercial-use conditions.
Plans
Every plan and priceEnterprise
Free
Team
Open source
The verdict
Surya offers document-focused recognition with structured outputs and interfaces for both command-line and interactive use. Before adopting it, account for the required inference backend and review the model-weight license conditions for your use case.
What the maker publishes
- Headquarters
- Manhattan, New York, United States source
- Inference backends
- Layout, OCR, and table recognition use vLLM on NVIDIA GPUs or llama.cpp on CPU and Apple Silicon, and can connect to an existing server. source
- Supported inputs
- both source
- Installation
- The README installs Surya with pip install surya-ocr and lists Docker plus NVIDIA Container Toolkit for NVIDIA GPU use, or llama.cpp for CPU and Apple Silicon. source
- Intended users
- Datalab says its platform serves AI/ML teams, enterprises, and product teams building AI agents, RAG systems, and document automation workflows. source
- Interfaces
- The project provides a Python package, command-line tools, and an interactive Streamlit app for images and PDFs. source
- Supported languages
- 91 source
- Languages
- The repository describes Surya as supporting OCR in 90+ languages. source
- License
- The Surya code is Apache 2.0, while the model weights use a modified AI Pubs Open Rail-M license with commercial-use conditions described in the README. source
- Limitations
- Surya is specialized for document OCR rather than photos or natural scenes, and layout, OCR, and table recognition require a running inference backend. source
- Managed access
- Datalab offers a managed platform running Surya and Chandra, with $5 in free credits and a free public playground. source
- Math
- Surya 2 recognizes equations inline in page OCR and returns them in KaTeX-compatible LaTeX inside math tags. source
- Output
- Surya 2 returns page content as structured blocks with HTML, polygons, bounding boxes, confidence, and reading order. source
- Primary platform
- desktop source
- Purpose
- Surya is a 650M-parameter OCR model for text recognition, layout analysis, reading order, and table recognition. source
- Security and compliance
- Datalab's pricing page lists a SOC 2 Type II report for every plan; BAA/DPA support is listed for Team and Enterprise. source
- Support
- Datalab lists community support on Free, email and Slack support on Team, and dedicated support on Enterprise. source
- Table recognition
- Table recognition can return row and column geometry and, in full mode, HTML with spanning cells and header rows. source
Checked against the maker's pages on 9 Oct 2026.