Atlas

Benchmarks

← All benchmarks

MMBench v1.1 (English, split unstated)

Multimodal · 2023-07-12

MMBench v1.1 English overall accuracy (CircularEval over 20 ability dimensions) from sources that do not say whether the dev or the test split was scored, such as the Qwen3-VL technical report and the Kimi-VL and Nemotron Nano V2 VL cards. Because the split is unknown it is kept as a diagnostic; results that name the dev or test split use mmbench-v1-1-dev-en or mmbench-v1-1-en-test. Higher is better.

Top models (higher is better)

ModelScore
Gemini 2.5 Pro90.1
Qwen3 VL 32B Thinking89.5
Qwen3-VL-235B-A22B-Instruct89.3
Qwen3 VL 235B A22B Thinking88.8
Qwen2.5-VL-72B-Instruct88.3
Qwen3 VL 32B Instruct87.6
Gemini 2.5 Flash87.1
Qwen3 VL 30B A3B Thinking87.0
GPT-5 Mini86.6
Qwen3 VL 30B A3B Instruct86.1
Qwen3 VL 4B Thinking84.6
Kimi-VL-A3B-Thinking-250684.4
Qwen3 VL 4B Instruct83.9
GPT-583.8
Nemotron Nano 12B V2 VL BF1683.0
Opus 4.183.0
Qwen3 VL 2B Thinking79.9
Gemma 3 27B IT78.9
Qwen3 VL 2B Instruct78.4
GPT-5 Nano78.4
GPT-4o Mini77.1
Kimi-VL-A3B-Thinking76.0
Gemma 3 12B IT74.6
Llama 3.2 11B Vision Instruct65.8
Loading Atlas data…