Atlas

Benchmarks

← All benchmarks

InfographicVQA

Multimodal · 2021-04-26

InfographicVQA asks questions about infographic images that combine text, layout, charts, icons and data visualizations, many of which require simple arithmetic or counting. This row is the 3,288-question test split scored by Average Normalized Levenshtein Similarity (ANLS). Higher is better.

Top models (higher is better)

ModelScore
Gemma 4 31B IT92.0
Qwen3 VL 235B A22B Thinking89.5
Gemma 4 26B A4B IT89.3
Qwen3-VL-235B-A22B-Instruct89.2
Qwen3 VL 32B Thinking89.2
Gemma 4 12B IT88.4
Qwen2.5-VL-72B-Instruct87.3
Qwen3 VL 32B Instruct87.0
Qwen3 VL 8B Thinking86.0
Qwen2 VL 72B Instruct84.5
Gemini 2.5 Pro84.2
InternVL2.5-78B84.1
Qwen2.5 VL 32B Instruct83.4
Qwen3-VL-8B-Instruct83.1
Qwen3 VL 4B Thinking83.0
Qwen2.5-VL-7B-Instruct82.6
Gemini 2.5 Flash82.5
Gemini 1.5 Pro (May 2024)81.0
Qwen3 VL 4B Instruct80.3
GPT-579.0
Qwen2.5-VL-3B-Instruct77.1
Qwen3 VL 2B Thinking77.1
Qwen2 VL 7B Instruct76.5
Gemini 1.5 Flash (May 2024 report)75.3
Gemini 1.5 Pro (February 2024 report)72.7
Qwen3 VL 2B Instruct72.4
Gemma 3 27B IT70.6
Gemma 4 E4B IT70.0
Opus 4.169.4
GPT-5 Nano68.6
Qwen2 VL 2B Instruct65.5
Gemma 3 12B IT64.9
Gemma 4 E2B IT63.9
Sonnet 458.0
Gemini Nano 254.5
Gemini Nano 151.1
Gemma 3 4B IT50.0
Loading Atlas data…