SoMark open-sources OXR, a 1.3B document parsing model that tops three benchmarks
AISoMark has open-sourced OXR, a document parsing model with 1.3B main-model parameters. The team reports first place on OmniDocBench, olmOCR-Bench and DocParsingBench, and average parsing time of about 101 seconds for 100 pages versus about 145 and 255 seconds for PaddleOCR-VL-1.6 and MinerU2.5-Pro on one RTX PRO 6000 Blackwell GPU. Code, weights, an online demo and a free API are available.
AIWhy it matters.
The article details OXR's benchmark scores, parsing speed comparison, and recursive self-improvement training loop, which show how a small document parser was built and measured.
Loading source posts…
Source details
QbitAI · 量子位的朋友们 · qbitai.comPublished · added here
Original title: OXR强势开源！1.3B横扫全球三榜第一，速度快过Paddle、MinerU，RSI让它卷出新高度