benchmarks.site

Public leaderboard · OCR & document understanding

OmniDocBench

Tests how well AI can read a PDF page and turn it into clean text, including tables and math formulas. It matters for turning scanned or complex documents into usable data.

Leaderboard

Top 3 right now

  1. 1stOXR-1.097.03
  2. 2ndTeleOCR96.91
  3. 3rdOvisOCR296.47

What it measures

PDF page parsing across diverse document types, scored on text, tables, formulas, layout and reading order with edit-distance and structure metrics.

How to read the numbers

Overall document-parsing score (0-100, higher is better), averaging text accuracy, table accuracy and formula accuracy — higher is better.

Results

25 rows · higher is better (Overall)CSV
1OXR-1.0Specialized VLMs1.3B97.030.024496.9396.61
2TeleOCRSpecialized VLMs1.2B96.910.026796.589596.8183
3OvisOCR2Specialized VLMs0.8B96.470.026597.485494.5842
4MiniCPM-OCR 1.0Specialized VLMs1.3B96.440.031698.081494.3999
5PaddleOCR-VL-1.6Specialized VLMs0.9B96.340.032697.530494.7619
6MinerU2.5-ProSpecialized VLMs1.2B95.750.03697.4593.42
7GLM-OCRSpecialized VLMs0.9B95.220.04497.1892.83
8PaddleOCR-VL-1.5Specialized VLMs0.9B94.930.03896.8991.67
9PaddleOCR-VLSpecialized VLMs0.9B94.180.04095.9190.65
10Unlimited-OCRSpecialized VLMs3B94.000.039495.723490.21
11Youtu-ParsingSpecialized VLMs2.5B93.740.04493.6392.02
12Qianfan-OCRSpecialized VLMs4B93.900.0495.0890.53
13Ovis2.6-30B-A3BGeneral VLMs30B93.700.03595.1789.44
14Logics-Parsing-v2Specialized VLMs4B93.330.04195.6588.42
15ABot-OCRSpecialized VLMs2B93.300.03794.8688.69
16FireRed-OCRSpecialized VLMs2B93.260.03795.4488.04
17MinerU-2.5Specialized VLMs1.2B93.040.04595.7787.88
18Gemini 3 ProGeneral VLMs-92.910.06495.9989.15
19Gemini 3 FlashGeneral VLMs-92.620.06695.1689.29
20MonkeyOCR V2Specialized VLMs-91.800.063693.498588.2628
21dots.ocrSpecialized VLMs3B90.770.04889.9587.18
22OpenDoc-0.1BSpecialized VLMs0.1B90.670.04993.0283.88
23DeepSeek-OCR 2Specialized VLMs3B90.250.05091.8483.89
24HunyuanOCRSpecialized VLMs1B89.950.08887.6891.01
25Qwen3-VL-235BGeneral VLMs235B89.780.06392.5583.07

Note: End-to-end table, benchmark version v1.6_full, from repo README in the order listed (best first), top 25. Scores are mostly self-reported by model authors or the benchmark team.

Source: OpenDataLab / Shanghai AI Lab (OmniDocBench repo) · github.com · results as of 2026-11-12 · retrieved 2026-11-17 · license: not stated (dataset research-only, non-commercial per repo copyright statement)

Markdown · More OCR & document understanding