SoMark has open-sourced OXR, a roughly 1.3B-parameter document-parsing model that separates layout analysis from content recognition and reading-order prediction. It can process 13 types of document elements into Markdown or JSON, while its post-training pipeline uses recursive self-improvement to turn errors and uncertain samples from unlabeled documents into new training data. The team reported first-place scores on three benchmarks and a 101-second end-to-end parse time for 100 pages on one RTX PRO 6000, but these performance claims have not been independently verified.