Atlas

Benchmarks

← All benchmarks

OCRBench

Multimodal · 2023-05-13

OCRBench (v1) combines 1,000 manually checked question-answer pairs over text recognition, scene-text and document VQA, key-information extraction and handwritten mathematical expressions. The official score is the count of correct answers out of 1,000, and rows are stored on that 0-1,000 points scale (sources that print a percentage are multiplied by 10, as each row's notes say). Its successor OCRBench v2 is a separate benchmark.

Top models (higher is better)

ModelScore
Qwen3.5 397B A17B931
Kimi K2.5923
Qwen3.5 122B A10B921
Qwen3-VL-235B-A22B-Instruct920
Qwen3.5 Plus (2026-02-15)914
Qwen3.5-Omni-Plus913
Qwen3.5 35B A3B910
Gemini 3 Pro Preview904
Qwen3 VL 30B A3B Instruct903
Qwen3.6 35B A3B900
Qwen3-VL-8B-Instruct896
Qwen3 VL 32B Instruct895
Qwen3.5-27B894
Qwen3.6 27B894
Qwen3.5-9B892
Qwen3.5-Omni-Flash891
Qwen2.5-VL-72B-Instruct885
Qwen3 VL 4B Instruct881
Gemini 3.5 Flash-Lite879
Qwen2 VL 72B Instruct877
Qwen3 VL 235B A22B Thinking875
Kimi-VL-A3B-Thinking-2506869
Gemini 2.5 Pro866
Opus 4.5865
GLM-4.5V865
Qwen2.5-VL-7B-Instruct864
Kimi-VL-A3B-Thinking864
Gemini 2.5 Flash864
LFM3-VL-2B863
Gemma 4 31B IT861
Qwen3.5-4B860
Qwen3 VL 2B Instruct858
Nemotron Nano 12B V2 VL BF16856
Qwen3 VL 32B Thinking855
Nemotron Nano 12B V2 VL854
InternVL2.5-78B854
Qwen3.5-2B854
MiniCPM-o 2.6852
MiniCPM-V 2.6852
LFM2.5-VL-3B842
Loading Atlas data…