OXR 1.3B document parser ranks first on three OCR benchmarks, per Rohan Paul

Rohan Paul, in an X post, says OXR, a 1.3B open-source document parser, ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench.

The post says OXR was trained with recursive self-improvement (RSI) in post-training: a VLM evaluates its outputs, disagreements between rollouts flag uncertainty, and a separate clustering model groups the causes so targeted data can be generated for the next round.

The post also says OXR runs at 3.05 times the throughput of GLM-OCR on a single RTX PRO 6000, which it describes as the fastest open-source result in its speed comparison. These are the author's claims; the post does not link to independent evaluation results.