OmniDocBench v1.6 Full - Reading Order Edit: leaderboard

OmniDocBench v1.6 Full - Reading Order Edit — Native OmniDocBench PRIMARY caption explicitly v1.6_full,1651pages. Pinned primary READMEf133a71e; update prose also mentionsv1.7 but this caption names the actual cohort. Multi-Granularity Adaptive Matching, groundtruthunchanged. Overall=((1-TextEdit)*100+TEDS+FormulaCDM)/3. Native primary cells retain fourdecimal precision; source OXRownresult extends exact v1.6_full cohort, with its README separately hashguarded. OXRcard peer copies not added as independent measurements. No fusion with existing v1.5 or unversioned othercohorts. Text/readingorder edit values remain native ratios, not percentages. No inferred model decoding settings. Metric: Normalized edit distance. Source: https://github.com/opendatalab/OmniDocBench

Metric: Normalized edit distance. Source: github.com. Saturation forecast: Around May 2027. 36 models tracked. Updated 16 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1OvisOCR20.112
2Youtu-Parsing0.116
3TeleOCR0.1184
4MinerU2.5-Pro0.12
5OXR-1.00.1214
6PaddleOCR-VL-1.60.1278
7Unlimited-OCR0.1281
8PaddleOCR-VL-1.50.13
9Qianfan-OCR0.13
10MinerU-2.5omnidocbench_v16_primary_with_oxr run0.13
11FireRed-OCR0.131
12GLM-OCR0.133
13PaddleOCR-VL0.135
14Ovis2.6-30B-A3B0.135
15Logics-Parsing-v20.137

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=omnidocbench-v1-6-full-reading-order-edit · How It Works · Data refreshed daily, snapshot 2026-11-16.