Surya

Surya is a 650M-parameter OCR system built for document text, layout, reading order, tables, and equations. It best suits developers and document-automation teams that can operate a local or hosted inference backend and need multilingual, structured results. It is a strong open-source option for page-centric workflows, provided you plan backend deployment and the model-weight license separately from the code license.

Overview

Surya targets documents rather than general scene text, supporting 91 languages for OCR, layout analysis, reading-order detection, and table recognition. Surya 2 emits structured page blocks with HTML, polygons, bounding boxes, confidence values, and reading order. Those fields are useful when a pipeline must preserve where content appeared and how it is organized, not merely produce a text string. Search indexing, retrieval-augmented generation, document automation, and agent workflows can consume the geometry and hierarchy directly.

Handwriting recognition and inline equation recognition broaden its document coverage. Equations are returned as KaTeX-compatible LaTeX inside math tags. Table recognition can return row and column geometry; full mode adds HTML with spanning cells and header rows. That makes Surya more suitable for forms, reports, and technical pages than a text-only OCR engine, but its document specialization makes it a poor fit for photographs or natural scenes.

Key features

Structured extraction

Coordinates, confidence, reading order, and HTML give downstream software the material to rebuild a page or apply field-level processing. Table geometry and spanning-cell HTML help when tabular meaning matters, while math tags avoid reducing equations to damaged plain text.

Interfaces and deployment

The Python package supports application integration, command-line tools suit repeatable batch jobs, and the Streamlit app offers an interactive way to inspect images and PDFs. Layout, OCR, and table recognition require a running inference backend: vLLM on NVIDIA GPUs, or llama.cpp on CPU and Apple Silicon, including connection to an existing server. Installation uses pip install surya-ocr; Docker and NVIDIA Container Toolkit are documented for NVIDIA deployments. This flexibility helps teams choose hardware, but it also makes operations part of the project rather than an invisible service.

Licensing

The code is Apache 2.0. Model weights use a modified AI Pubs Open Rail-M license with commercial-use conditions described in the project documentation, so commercial deployment needs a separate weights-license review.

Pricing

The software is free and open source, with no charge for the self-hosted package. Datalab also offers managed access to Surya and Chandra, including $5 in free credits and a free public playground.

Free

The managed Free plan costs nothing, billed at $0 pay-as-you-go. It includes a $20 monthly allowance with a work email or $10 with a personal email, 25 requests per minute, 25 concurrent requests, and community support. It is appropriate for evaluation and small workloads; the allowance and rate limits are unlikely to suit a busy production service.

Team

Team costs $400 a month, billed $400 monthly, with $400 of usage included and pay-as-you-go thereafter. Capacity rises to 400 requests per minute and 400 concurrency, with email and Slack support. This fits product teams that need materially higher throughput and responsive support, but the fixed monthly commitment is substantial compared with self-hosting.

Enterprise

Enterprise uses custom pricing and custom usage, rate limits, and terms. It adds dedicated support plus VPC or fully air-gapped deployment, making it the relevant option for organizations with isolation or procurement requirements.

Datalab lists SOC 2 Type II for every plan. BAA and DPA support are listed for Team and Enterprise, which matters for regulated deployments.

Platforms

Surya supports API, Linux, macOS, and self-hosted operation, with desktop as its primary platform. CPU and Apple Silicon users can use llama.cpp; NVIDIA GPU users can use vLLM. The package accepts both supported input types, and the interactive app handles images and PDFs.

Who it's for

Choose Surya for AI/ML teams, enterprises, and product groups building RAG systems, agents, or document-automation workflows that need coordinates, hierarchy, tables, handwriting, or equations across many languages. It is especially compelling when self-hosting or connecting to an existing inference server is acceptable. Look elsewhere if you need turnkey OCR for photographs, do not want to operate an inference backend, or require a simpler license review for commercial weights.

Pros and cons

Pros

• Rich structured output: HTML, geometry, confidence, and reading order support reliable downstream document processing.
• Broad document coverage: 91 languages, handwriting, tables, and KaTeX-compatible equations address more than basic printed text.
• Flexible interfaces: Python, command line, and Streamlit cover integration, automation, and visual inspection.
• Deployment choice: vLLM, llama.cpp, or an existing server supports NVIDIA, CPU, and Apple Silicon environments.

Cons

• Backend overhead: Core recognition functions need a running inference backend, adding infrastructure and maintenance work.
• Scope limitation: It is specialized for documents, not photos or natural scenes.
• Weights licensing: Commercial-use conditions differ from the Apache-licensed code and require review.
• Managed Free limits: Allowances, 25 requests per minute, and 25 concurrency favor testing over sustained production traffic.

Alternatives

OCR.space is a better fit when you want a freemium hosted/API option with broad platform coverage; its Free plan allows 25,000 requests per month but limits files to 1 MB, PDFs to three pages, and adds a searchable-PDF watermark.

Tesseract OCR suits users seeking a free, Apache 2.0 open-source OCR engine without Surya's structured document features.

docTR is the alternative for a free open-source Python library installable with pip, Git, or Docker on Python 3.11 or higher.

EasyOCR is another free self-hosted OCR option for Linux, macOS, and Windows.

OCRmyPDF is preferable when the job is self-hosted OCR processing for PDFs and you are comfortable with external OCR and PDF-tool dependencies.

PaddleOCR offers an official API free tier with a daily document-parsing limit of 20,000 pages.

Veryfi Invoices OCR API is the specialized choice for invoice workflows, with a Starter plan at $500 a month and per-invoice pricing.

Amazon Textract AnalyzeExpense fits teams already using AWS that need expense-document analysis through an API, with a three-month new-customer free tier covering up to 100 pages per month.

Browse the OCR API Software and OCR software lists for more options.

Verdict

Surya is a strong choice for developers building multilingual document pipelines where layout, tables, equations, handwriting, and coordinates matter. Its main reason to choose it is the depth of structured output across flexible self-hosted interfaces. Look elsewhere if backend operations, document-only scope, or the model-weight commercial terms do not fit your project.