Description
OXR-1.0 is an advanced open-source document parsing model developed by the SoMark team, designed to convert complex documents such as PDFs and images into structured, agent-friendly data. It outputs data in formats like Markdown or JSON, complete with layout coordinates, making it highly versatile for various applications requiring precise document understanding.
Read More
Distinguished by its innovative architecture, OXR-1.0 is the first document parsing model to incorporate a recursive self-improvement (RSI) loop during post-training. This unique approach allows the model to enhance its performance by targeting samples it processes incorrectly or weakly, significantly accelerating iteration and improvement without requiring human annotation. Despite having only 1.3 billion parameters, OXR-1.0 consistently ranks #1 on major benchmarks including olmOCR-Bench, OmniDocBench, and DocParsingBench, demonstrating superior accuracy, particularly in table parsing.
A key feature of OXR-1.0 is its pluggable layout analysis module, a first in the industry. This design allows users to replace the built-in layout model with a custom one, providing unparalleled flexibility to adapt to diverse domain-specific requirements. OXR-1.0 maintains its core capabilities in content recognition and reading order, ensuring minimal integration effort across various document scenarios. It can recognize 13 distinct document element types, from formulas and tables to chemical equations and code blocks, providing bounding box coordinates for each.
Beyond its accuracy, OXR-1.0 boasts top-tier parsing speed, outperforming other open-source models in speed tests. This efficiency is achieved through a redesigned processing architecture at the task level, optimizing speed while preserving high accuracy. The model is capable of handling challenging document types and layouts, including A3 exam papers, newspapers, vertical text, and even handwriting. SoMark provides a free API for testing and offers comprehensive CLI and SDK tools for Python and JavaScript, ensuring broad compatibility and ease of integration for developers.
- CategoryAI Models & LLMs
- Pricing ModelFree API / Open Source
- DeploymentCloud, Runs locally
- PlatformsWeb
OXR-1.0 Highlights
Open-source document parsing model
Converts PDFs and document images to Markdown or JSON
Includes layout coordinates in output
Utilizes Recursive Self-Improvement (RSI) loop for post-training
Achieves State-of-the-Art (SOTA) accuracy
Ranks #1 on olmOCR-Bench, OmniDocBench, and DocParsingBench
Features 1.3 billion parameters
Offers industry-leading parsing speed
Independent and pluggable layout analysis module
Supports custom layout models
Handles content recognition and reading order
Recognizes 13 document element types (formulas, tables, chemical equations, code blocks, etc.)
Performs well on challenging documents (A3 exam papers, handwriting, vertical text)
Provides a free API for testing
Compatible with SoMark's official document parsing API and CLI tools
Getting Started with OXR-1.0
Access Model: Download the OXR-1.0 code and models from Hugging Face or ModelScope.
Prepare Environment: Set up your environment for model download and service execution.
Start Services: Run OXR-1.0 services using Docker for a recommended setup, or start individual services (OXR Model, Layout Analysis Model, Document Parsing API) from the source.
Configure API: If running from source, configure the Document Parsing API by setting URL fields in config.yaml to the addresses of the OXR and Layout Analysis model services.
Open Visual Demo: Access the visual demo at http://localhost:8000/demo to upload PDFs or images and view parsed results, including Markdown preview, source, layout bounding boxes, and reading order.
Send Parsing Request: Use the OXR API directly or integrate with SoMark's CLI/SDK tools (somark-py, somark-js) by changing the Base URL to send parsing requests.
Optimize Integration: Adapt the pluggable layout analysis module to specific domain requirements for customized document structure parsing.
OXR-1.0's Use Cases
- Document Digitization
- Data Extraction
- Content Recognition
- Automated Document Processing
- Agent-Friendly Data Preparation
- Custom Layout Analysis
- Research and Development
- Benchmarking Document Parsers








