Skip to content

Benchmark

MMBench EN

Vision benchmarks measure how well models understand images. Scores are percentages.

Models measured
39
Scores recorded
39
Sources
1
IQ weight
0.93

Score distribution

76.5%
Mean
79.3%
Median
14.2%
Std dev
67.8%
10th percentile
86.9%
90th percentile
1.1%
Min
1.1%88.8%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.93 for chat models, measured over 39 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1InternVL3 78B· OpenVLM Leaderboard data88.8%
2Qwen2.5 VL 72B Instruct· OpenVLM Leaderboard data88.3%
3internvl3-38b· OpenVLM Leaderboard data87.7%
4GPT 4.1· OpenVLM Leaderboard data86.9%
5GPT-4.1 mini· OpenVLM Leaderboard data86.9%
6Qwen2-VL-72B· OpenVLM Leaderboard data86.9%
7Hunyuan Standard Vision· OpenVLM Leaderboard data86.4%
8Grok 2 Vision 1212· OpenVLM Leaderboard data85.3%
9DeepSeek VL2· OpenVLM Leaderboard data84.1%
10internvl3-8b· OpenVLM Leaderboard data83.6%
11InternVL2 26B· OpenVLM Leaderboard data83.4%
12Phi 4 Multimodal Instruct· OpenVLM Leaderboard data83.4%
13Qwen2.5-VL 7B Instruct· OpenVLM Leaderboard data83.2%
14Gemma 3 27B IT· OpenVLM Leaderboard data82.3%
15MiniCPM-V 2.6· OpenVLM Leaderboard data81.5%
16GLM 4 Vision Plus· OpenVLM Leaderboard data81.4%
17GLM 4 Vision· OpenVLM Leaderboard data81.3%
18Llama 3.2 90B Vision Instruct· OpenVLM Leaderboard data80.4%
19Gemini 1.5 Flash· OpenVLM Leaderboard data79.4%
20Moonshot V1 8K· OpenVLM Leaderboard data79.3%
21qwen2.5-vl-3b-instruct· OpenVLM Leaderboard data79.1%
22InternVL2 4B· OpenVLM Leaderboard data78.5%
23Yi Vision· OpenVLM Leaderboard data78.1%
24Pixtral 12B· OpenVLM Leaderboard data77.9%
25Qwen-VL Max· OpenVLM Leaderboard data77.6%
26Gemma 3 12B IT· OpenVLM Leaderboard data77.5%
27Phi-3.5 Vision Instruct· OpenVLM Leaderboard data76%
28Gemini 1.5 Pro· OpenVLM Leaderboard data73.9%
29gemini-1.0-pro· OpenVLM Leaderboard data73.6%
30Molmo 7B-D· OpenVLM Leaderboard data73.6%
31Moondream2· OpenVLM Leaderboard data70%
32Gemma 3 4B· OpenVLM Leaderboard data69.6%
33GPT-4.1 nano· OpenVLM Leaderboard data68.2%
34Llama 3.2 11B Vision Instruct· OpenVLM Leaderboard data68%
35Claude 3 Sonnet· OpenVLM Leaderboard data67.8%
36Qwen-VL Plus· OpenVLM Leaderboard data67%
37Claude 3 Opus· OpenVLM Leaderboard data63.3%
38Claude 3 Haiku· OpenVLM Leaderboard data60.7%
39kosmos-2· OpenVLM Leaderboard data1.1%

Data credits: OpenVLM Leaderboard (OpenCompass / VLMEvalKit), Apache-2.0 (licence)