OXR – An RSI-trained 1.3B open-source SOTA document parsing model (github.com)

The SoMark team has introduced OXR (Optical Everything Recognition), an innovative open-source document parsing model featuring state-of-the-art (SOTA) accuracy and speed. OXR, with 1.3 billion parameters, transforms PDFs and document images into structured, Agent-Friendly data formats like Markdown and JSON, while also incorporating layout coordinates. Uniquely, it employs a recursive self-improvement (RSI) loop during post-training, enhancing its capabilities by autonomously learning from its past mistakes, a groundbreaking approach in document parsing. OXR achieves top rankings on multiple benchmarks, demonstrating significant advantages in parsing speed and accuracy, particularly for complex layouts such as tables and handwriting. Significantly, OXR offers a pluggable layout analysis module, allowing users to customize how document structures are analyzed without sacrificing recognition capabilities. This flexibility enables seamless adaptation to varying document scenarios and business needs, making it a versatile tool for developers. The system recognizes 13 different document element types, reinforcing its utility across diverse applications, from academic papers to technical documents. The free API provided with OXR allows users to test and integrate the model easily, marking a step forward in accessibility for advanced document parsing technology.