Turn documents into Agent-Friendly data! OXR is a 1.3B open-source document parsing model that turns PDFs and document images into structured Markdown or JSON. Instead of extracting text alone, it recognizes 13 different document elements including tables, formulas, figures, code blocks, footnotes, and chemical structures, while preserving reading order and layout coordinates. That makes the output much more useful for RAG and agents, where knowing what the text says is only one part of the problem. You also need to know what is a table, where a formula belongs, which caption maps to which figure, and how the document should actually be read. The technical part I found interesting is how OXR was post-trained. It uses a recursive self-improvement loop where OXR first parses unlabeled documents, then a general VLM reviews the outputs for errors and uncertainty. Disagreement across multiple OXR rollouts helps surface uncertain examples, those failures are grouped by cause, and the VLM generates targeted training data for the next round. So instead of manually labeling new data for every iteration, the model’s own parsing failures help decide what it should train on next. The architecture is also different from the usual parsing pipeline. Layout analysis stays independent and pluggable, while recognition and content-aware reading order run in parallel inside the main model. That means teams can replace the layout model for their own document structure without replacing the rest of OXR. And the results are strong for a 1.3B model. OXR ranks #1 on olmOCR-Bench, OmniDocBench, and DocParsingBench. In the reported 100-page speed test on a single RTX PRO 6000, OXR finished in about 101 seconds vs. 161 seconds for DeepSeek-OCR-2. The full model is open source.