DocParsingBench: leaderboard

Systems turn 1,400 pages of real business documents (finance, legal, research, manufacturing, education; Chinese and English) into Markdown; the README does not state how many pages each system was scored on. The score is DPB, SoMark's overall mark built from text edit distance, formula CDM and table TEDS with weights 0.5, 0.3 and 0.2. Unlike OmniDocBench, the pages come from production documents with stamps, scans and handwriting and the order of segments is matched rather than fixed.

Metric: DPB overall score (0-1). Source: github.com. Saturation forecast: Rough model projection: around 2026. 17 models tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1OXR-1.00.8776
2PaddleOCR-VL-1.60.8517
3Jina-OCR-v10.8446
4FireRed-OCR0.8405
5TeleOCR0.8313
6Dots.mocr0.8306
7MinerU2.5-Pro0.8278
8Infinity-Parser2-Pro0.8263
9OvisOCR20.8226
10Chandra-OCR-20.822
11Qianfan-OCR0.8145
12ABot-OCR0.8042
13GLM-OCR0.8034
14Youtu-Parsing0.7625
15HunyuanOCR 1.50.7535

Interactive version: theaggregate.ai/benchmark?slug=docparsingbench · How It Works · Data refreshed daily, snapshot 2026-11-16.