OCRBench

AI benchmark results and model performance

Type: percentageImage To Text

Recorded Results

Average Score

559%

Score Range

80.90% - 920%

Leading Model (lowest score)

80.90% - DeepSeek VL2 Tiny

Scores Over Time

Individual benchmark scores plotted by date.

Models Using This Benchmark

Organisation	Model	Reported	Top Score	Info	Self Reported	Source
DeepSeek	DeepSeek VL2 Tiny	13 Dec 2024	80.90%	-	Yes	Source
DeepSeek	DeepSeek VL2	13 Dec 2024	81.10%	-	Yes	Source
DeepSeek	DeepSeek VL2 Small	13 Dec 2024	83.40%	-	Yes	Source
Microsoft	Phi 4 multimodal instruct	01 Feb 2025	84.40%	-	Yes	Source
Qwen	Qwen 2.5 0.5B	-	86.40%	inferred family alias from qwen2.5-vl-7b (score=0.3131; benches=32)	Yes	Source
Qwen	Qwen 2.5 1.5B	-	86.40%	inferred family alias from qwen2.5-vl-7b (score=0.3131; benches=32)	Yes	Source
Qwen	Qwen 2.5 3B	-	86.40%	inferred family alias from qwen2.5-vl-7b (score=0.3531; benches=32)	Yes	Source
Qwen	Qwen 2.5 VL 7B Instruct	-	86.40%	-	Yes	Source
Qwen	Qwen 2 VL 2B	-	87.70%	inferred family alias from qwen2-vl-72b (score=0.4083; benches=15)	Yes	Source
Qwen	Qwen 2 VL 72B	-	87.70%	-	Yes	Source
Qwen	Qwen 2.5 Math 72B	-	88.50%	inferred high-confidence family alias from qwen2.5-vl-72b (score=0.4700; benches=30)	Yes	Source
Qwen	Qwen 2.5 Math PRM 72B	-	88.50%	inferred family alias from qwen2.5-vl-72b (score=0.4050; benches=30)	Yes	Source
Qwen	Qwen 2.5 Math RM 72B	-	88.50%	inferred family alias from qwen2.5-vl-72b (score=0.4092; benches=30)	Yes	Source
Qwen	Qwen 2.5 VL 72B Instruct	26 Jan 2025	88.50%	-	Yes	Source
Qwen	Qwen 3.5 27B	24 Feb 2026	89.40%	Raw score: 89.4	Yes	Source
Qwen	Qwen 3.5 Flash	23 Feb 2026	89.40%	inferred family alias from qwen3.5-27b (score=0.4147; benches=81)	Yes	Source
Qwen	Qwen 3.5 35B A3B	24 Feb 2026	91%	Raw score: 91.0	Yes	Source
Qwen	Qwen 3.5 122B A10B	24 Feb 2026	92.10%	Raw score: 92.1	Yes	Source
Moonshot	Kimi K2.5	27 Jan 2026	92.30%	-	Yes	Source
Qwen	Qwen 3 4B Instruct 2507	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 4B SafeRL	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 4B Thinking 2507	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3462; benches=48)	Yes	-
Qwen	Qwen 3 Embedding 4B	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 Guard Gen 4B	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3400; benches=48)	Yes	-
Qwen	Qwen 3 Guard Stream 4B	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3371; benches=48)	Yes	-
Qwen	Qwen 3 Reranker 4B	-	808%	inferred family alias from qwen3-vl-4b-thinking (score=0.3850; benches=48)	Yes	-
Qwen	Qwen 3 VL 4B Thinking	-	808%	-	Yes	-
Qwen	Qwen 3 8B	-	819%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.4600; benches=50)	Yes	-
Qwen	Qwen 3 Guard Stream 8B	-	819%	inferred family alias from qwen3-vl-8b-thinking (score=0.3371; benches=50)	Yes	-
Qwen	Qwen 3 VL 8B Thinking	-	819%	-	Yes	-
Qwen	Qwen 3 VL Embedding 8B	-	819%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5232; benches=50)	Yes	-
Qwen	Qwen 3 VL Reranker 8B	-	819%	inferred high-confidence family alias from qwen3-vl-8b-thinking (score=0.5275; benches=50)	Yes	-
Qwen	Qwen 3 VL 30B A3B Thinking	-	839%	-	Yes	-
Qwen	Qwen 3 VL 32B Thinking	-	855%	-	Yes	-
Qwen	Qwen 3 VL 235B A22B Thinking	-	875%	-	Yes	-
Qwen	Qwen 3 VL 4B Instruct	-	881%	-	Yes	-
Qwen	Qwen 3 VL 32B Instruct	-	895%	-	Yes	-
Qwen	Qwen 3 VL 8B Instruct	-	896%	-	Yes	-
Qwen	Qwen 3 VL 30B A3B Instruct	-	903%	-	Yes	-
Qwen	Qwen 3 VL 235B A22B Instruct	-	920%	-	Yes	-