OCRBenchMultimodal

OCRBench was introduced in 2023 by South China University of Technology and Huazhong University of Science and Technology. It combines 29 datasets across text recognition, scene-text VQA, document VQA, key-information extraction, and handwritten-formula recognition, using a unified 1,000-point score.

Ranking Sources

#ModelScore

1

Qwen3.5-397B-A17BAlibaba

93.1%

2

Gemini 3 ProGoogle

90.4%

3

Qwen3.6-35B-A3BAlibaba

90.0%

4

Qwen3.6-27BAlibaba

89.4%

5

Claude Opus 4.5Anthropic

85.8%

6

GPT-5.2OpenAI

80.7%

6 models total

Back to benchmarks