SoMark open-sources OXR: a 1.3B document parser that tops three benchmarks, with pluggable layout analysis
Synopsis
The SoMark team has open-sourced OXR, a document parsing model that turns PDFs and document images into Markdown or JSON with layout coordinates for AI agents. Pretrained from scratch, OXR adds a recursive self-improvement (RSI) loop during post-training that targets samples the model gets wrong or handles weakly, without human annotation. SoMark reports that OXR, at 1.3B parameters, ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench, and first among open-source document parsing models in speed tests on an RTX PRO 6000.
Interpretation
The open-source document parsing model OXR turns PDFs and document images into Markdown or JSON with layout coordinates, output that AI agents can use directly. Earlier document parsing tools either assigned layout analysis to the recognition model or used a dedicated model that also handled reading order. SoMark reports that OXR, at 1.3B parameters, ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench, with a particularly wide lead in table parsing.
OXR makes layout analysis an independent module whose built-in layout model can be replaced, while OXR still handles recognition and reading order. Layout analysis varies more across document domains than any other module and is typically the part document parsing providers need to customize most. The system consists of a layout analysis model and OXR, with reading order folded into the recognition stage and handled by the same model; SoMark says any layout analysis model can work with OXR.
A recursive self-improvement loop in post-training keeps training on samples the model gets wrong or handles weakly, without human annotation. SoMark says this loop made iteration faster than before and calls it a first in document parsing. OXR is pretrained from scratch and recognizes 13 document element types, including formulas, tables, chemical equations, chemical structures and code blocks, with bounding box coordinates.
Perspective
The demonstrated scope is parsing of PDFs and document images into Markdown or JSON with layout coordinates, covering layouts such as A3 exam papers, newspapers, vertical text and handwriting; local runs require Linux, an NVIDIA GPU with BF16 support and at least 24GB of GPU memory.
The three-benchmark rankings and speed conclusions currently come from SoMark's own description, and the technical report is marked as coming soon; the free API does not guarantee service stability, requests are queued and processing speed may vary, and SoMark does not recommend it for production.
