OXR, an Open-Weight 1.3B Document Parser, Claims Top Spots on Three Public Benchmarks

SoMarkAI released OXR (Optical Everything Recognition), a 1.3-billion-parameter document parsing model pretrained from scratch, with code and weights published on 20 October. The model converts document PDFs and images into Markdown or JSON with layout coordinates and reading order, recognises 13 element types including tables, formulas, chemical equations and code blocks, and handles difficult inputs such as A3 exam papers, newspapers, vertical text and handwriting. It claims first place on olmOCR-Bench, OmniDocBench and DocParsingBench, plus the fastest end-to-end time in a 21-model comparison, and claims a first in document parsing: recursive self-improvement during post-training, where uncertain samples are clustered by failure cause to plan the next training round.

The layout module is independently pluggable, serving runs on vLLM, and an all-in-one Docker image plus a free hosted API are available for evaluation. The licence is Apache 2.0-based but commercial use is free only for organisations under $2 million in total funding or annual revenue, with separate commercial terms above that threshold. Benchmark figures are vendor-driven.