OCR is getting this competitive too. Xiaomi just shared Xiaomi-OCR-0, a 0.8B o… · @Aiyafun · TwKit

@Aiyafun

OCR is getting this competitive too. Xiaomi just shared Xiaomi-OCR-0, a 0.8B open model under Apache-2.0 that puts document parsing, key information extraction and OCR question answering into one model, and the first thing I wanted to compare it with was Baidu's PaddleOCR-VL-1.6, the 0.9B parser they open-sourced in May, which has been my reference point for small document models. Xiaomi's technical report actually puts both in the same tables, with Xiaomi-OCR-0 at 96.83 overall on OmniDocBench v1.6 against 96.33 for PaddleOCR-VL-1.6, and the gap gets wider on Real5-OmniDocBench, where pages are rescanned, warped, skewed or shot under bad lighting, at 95.24 versus 93.19. These are Xiaomi's own runs though, Xiaomi-OCR-0 isn't on the official OmniDocBench leaderboard yet, where PaddleOCR-VL-1.6 sits at 96.34, and Xiaomi's pipeline uses Baidu's PP-DocLayoutV3 for layout detection, so to me it reads less like one replacing the other and more like the two building on each other's work. If you parse invoices, scanned contracts or phone photos of paper in real projects, what actually decides it for you, a point or two on a benchmark, or how the model handles your own messiest pages?

曝光 30 · 评论 0 · 点赞 2 · 书签 0