PureDocBench Leaderboard

Search or filter models, then click any metric header to sort. The first click follows the metric direction: descending for ↑ and ascending for ↓.

README leaderboard →
Overall / FormulaCDM / TableTEDS / Avg₃: higher is better · TextEdit / ROEdit: lower is better
Rank Architecture Params Clean Digital Degraded Real Degraded

Model names link to an official GitHub repository. Kimi K2.6 links to the Kimi K2 family repository; Gemini 3.1 Pro links to Google's Gen AI SDK. Scores for the original 40 models reproduce Table 2 of the PureDocBench paper. OvisOCR2 and NaviDC-OCR are author-reported post-publication results from their official model card and repository, respectively. Unreported component metrics are shown as —.