Description
Surya OCR 2.5 is a state-of-the-art image-text-to-text model developed by Datalab, specifically engineered for robust document intelligence tasks. With 689 million parameters, it excels in optical character recognition (OCR), layout analysis, and table recognition, offering significant improvements over previous versions.
Read More
The model demonstrates exceptional accuracy, scoring 85.9 on OmniParseBench and 83.7% on olmOCR-bench, making it the highest-scoring open model under 1 billion parameters in Datalab's evaluations. It boasts impressive speed, achieving a throughput of 4.5 pages per second on an H100 GPU at 192 DPI, which is 1.7 times faster than Surya 2 with MTP.
Surya OCR 2.5 is highly multilingual, achieving a 74.9% pass rate on an internal benchmark covering 91 languages. It performs comprehensive layout analysis, identifying elements such as tables, images, and headers, and accurately determines reading order. A key feature is its ability to output tables as HTML, complete with merged cells represented by `colspan` and `rowspan` attributes.
Beyond full-page OCR, Surya OCR 2.5 also includes smaller, specialized models for line-level text detection and OCR error detection. It supports a wide range of document types, from newspapers and reference books to government forms, handwritten letters, and corporate documents. The model can be installed and run with inference backends like vllm (for NVIDIA GPUs) or llama.cpp (for CPU/Apple Silicon), and an interactive Streamlit app is available for easy testing.
Commercial usage of the Surya code is licensed under Apache 2.0, while the model weights use a modified AI Pubs Open Rail-M license, allowing free use for research, personal projects, and startups under $5M funding/revenue. Broader commercial licensing options are available through Datalab's pricing page. Datalab also offers a managed platform that runs Surya and its higher-accuracy variant, Chandra, with free credits for new sign-ups.
- CategoryComputer Vision Tools
- Pricing ModelFreemium
- DeploymentCloud
- PlatformsWeb
Surya OCR 2.5 Highlights
Image-Text-to-Text Model: Processes images to extract text and structural information.
High Accuracy: Scores 85.9 on OmniParseBench and 83.7% on olmOCR-bench.
Fast Throughput: Achieves 4.5 pages/s on H100 at 192 DPI.
Multilingual Support: Benchmarked across 91 languages with 74.9% overall pass rate.
Layout Analysis: Detects tables, images, headers, and determines reading order.
Table Recognition: Extracts tables as HTML with merged cells (colspan/rowspan).
Text Line Detection: Separate model for precise line-level text detection.
OCR Error Detection: Dedicated model for identifying OCR inaccuracies.
VLM Integration: Runs layout, OCR, and table recognition through a single VLM.
Flexible Inference Backends: Supports vllm (NVIDIA GPU) and llama.cpp (CPU/Apple Silicon).
Interactive Streamlit App: Provides an interactive interface for testing images and PDFs.
Customizable Settings: Override settings via environment variables for fine-tuning.
Commercial Licensing: Apache 2.0 for code, modified Open Rail-M for model weights.
Managed Platform Access: Available on Datalab's managed platform with free credits.
Math/Equation Handling: Recognizes and outputs mathematical equations in KaTeX-compatible LaTeX within HTML.
Getting Started with Surya OCR 2.5
Install Surya OCR: Install the `surya-ocr` package using pip.
Set up Inference Backend: Configure either vllm (for NVIDIA GPU) or llama.cpp (for CPU/Apple Silicon) as the inference server.
Run OCR Command: Execute the OCR command with your image or PDF path to extract text and bounding boxes.
Perform Layout Analysis: Use the layout command to detect document structure and reading order.
Extract Tables: Utilize the table recognition command to get tables as HTML output.
Process Text Line Detection: Run the text line detection command to identify individual text lines.
Integrate via Python: Access Surya OCR functionalities directly from Python scripts.
Use Interactive App: Launch the Streamlit app to interactively test Surya on images or PDF files.
Surya OCR 2.5's Use Cases
- Document Digitization
- Data Extraction
- Content Archiving
- Multilingual Document Processing
- Layout Understanding
- Automated Form Processing
- Research and Academia
- Business Process Automation








