The Chinese startup team SoMark, specializing in intelligent document analysis, has made the OXR model, which stands for Optical Everything Recognition, publicly available. The main model, OXR-1.0, has only 1.3 billion parameters and is capable of converting PDF files and document images into Markdown or JSON formats, including element location coordinates, which the team describes as a data representation convenient for agents.
The code and model weights are available on GitHub, Hugging Face, and ModelScope platforms, and an online demo and free API are also provided.
Testing Results and Model Capabilities
According to the official repository, OXR ranks first in three public benchmarks: 97.18 points on OmniDocBench, 87.84 on olmOCR-Bench, and 87.76 on DocParsingBench, demonstrating a particularly significant advantage in table parsing. The team publishes full results and reproduction instructions in its benchmark documentation.
In a performance test conducted on one RTX PRO 6000 Blackwell card with 96 GB of memory, OXR processed 100 randomly selected pages from OmniDocBench in approximately 101 seconds. This was faster than PaddleOCR-VL-1.6 (approximately 145 seconds) and MinerU2.5-Pro (255 seconds).
The model recognizes 13 types of elements, including text, headings, tables, formulas, stamps, code blocks, chemical equations, and chemical structures. Tables are outputted in HTML or Markdown format, formulas in LaTeX, chemical equations in mhchem syntax, and molecular structures in SMILES strings. Furthermore, the model determines the reading order, which is critical for multi-column materials such as newspapers, scientific articles, and exam papers, where correctly recognized blocks might be assembled in the wrong sequence.
Innovative Architectural Solutions
Two key design decisions stand out. Firstly, a lightweight layout model is responsible for area detection, while the main model performs content recognition and reading order determination in a single step, eliminating a separate ordering stage. SoMark claims this is the first implementation that makes the layout module independent and pluggable, allowing businesses to replace their own layout module with other element categories or area boundaries while retaining OXR recognition.
Secondly, after training, a recursive self-improvement loop is used: OXR processes unlabeled documents, external models evaluate the resulting output and diagnose errors, a clustering model groups failures by cause, and target training data is generated for the next round without the need for manual labeling.
The team reported that the repository reached 1000 stars on GitHub in the first week. The publication of a scientific paper on arXiv is announced soon. For developers creating retrieval-augmented generation pipelines or agentic pipelines, OXR represents a small, fast, and open option for transforming complex documents into structured input.
