PureDocBench: leaderboard

PureDocBench (2026) scores document parsers and vision-language models on 1,475 pages in 10 domains rendered from freshly generated HTML/CSS, whose annotations come from the same source, presented clean, digitally degraded and physically recaptured. The score is the mean of the three track Overall scores, each averaging text accuracy, formula CDM and table TEDS on a 0 to 100 scale; higher is better.

Metric: Three-track average (0-100): mean of the clean, digitally degraded and real degraded track Overall scores, each the mean of text accuracy (100 times one minus the text normalized edit distance), formula CDM and table TEDS, on 1,475 source pages in 10 domains and 66 subcategories, rendered from LLM-written HTML/CSS with source-linked annotations for text, formulas, tables and reading order; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 58 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GLM-5.3 Flash79.69#137
2TeleOCR78.55
3Gemini 3.6 Flash78.27#84
4Qwen 3.8 27B77.63#232
5Claude Opus 577.07#26
6WeVisDoc-4B75.57
7OvisOCR275.1
8Qwen 3.5 122B A10B74.11#350
9FD-RL73.92
10WeVisDoc-2B73.69
11Logics-Parsing-v272.61
12Qwen 3.5 9B70.89#838
13Gemini 3.1 Pro (Preview)70.43#102
14Dots.mocr70.39
15ABot-OCR70.37

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=puredocbench · How It Works · Data refreshed daily, snapshot 2026-11-22.