Skip to content

Benchmark

LLaVA-Bench

Vision benchmarks measure how well models understand images.

Models measured
41
Scores recorded
41
Sources
1
IQ weight
0.86

Score distribution

82.8
Mean
85.2
Median
14.3
Std dev
66.5
10th percentile
99.6
90th percentile
33.6
Min
33.6105.6
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.86 for chat models, measured over 41 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1GPT 4.1· OpenVLM Leaderboard data105.6
2Gemma 3 27B IT· OpenVLM Leaderboard data103.1
3InternVL3 78B· OpenVLM Leaderboard data100.4
4Grok 2 Vision 1212· OpenVLM Leaderboard data100
5GPT-4.1 mini· OpenVLM Leaderboard data99.6
6Qwen2.5 VL 72B Instruct· OpenVLM Leaderboard data98.7
7Gemini 2.0 Flash· OpenVLM Leaderboard data97.7
8Gemini 1.5 Pro· OpenVLM Leaderboard data95.3
9internvl3-38b· OpenVLM Leaderboard data94.7
10InternVL2 26B· OpenVLM Leaderboard data93.1
11Gemma 3 12B IT· OpenVLM Leaderboard data93
12GLM 4 Vision Plus· OpenVLM Leaderboard data92.9
13Gemma 3 4B· OpenVLM Leaderboard data92.5
14Qwen2.5-VL 7B Instruct· OpenVLM Leaderboard data91
15DeepSeek VL2· OpenVLM Leaderboard data89.7
16GPT-4.1 nano· OpenVLM Leaderboard data88.8
17GLM 4 Vision· OpenVLM Leaderboard data87.5
18Llama 3.2 90B Vision Instruct· OpenVLM Leaderboard data86.5
19Llama 3.2 11B Vision Instruct· OpenVLM Leaderboard data85.8
20MiniCPM-V 2.6· OpenVLM Leaderboard data85.4
21internvl3-8b· OpenVLM Leaderboard data85.2
22Molmo 72B· OpenVLM Leaderboard data82.4
23Qwen-VL Max· OpenVLM Leaderboard data82.3
24Pixtral 12B· OpenVLM Leaderboard data81.5
25Hunyuan Standard Vision· OpenVLM Leaderboard data80.8
26Gemini 1.5 Flash· OpenVLM Leaderboard data80
27gemini-1.0-pro· OpenVLM Leaderboard data79.9
28Qwen2-VL-72B· OpenVLM Leaderboard data77.1
29qwen2.5-vl-3b-instruct· OpenVLM Leaderboard data77
30Moonshot V1 8K· OpenVLM Leaderboard data76.5
31Yi Vision· OpenVLM Leaderboard data76
32Claude 3 Opus· OpenVLM Leaderboard data73.9
33Qwen-VL Plus· OpenVLM Leaderboard data73.7
34Claude 3 Sonnet· OpenVLM Leaderboard data73.2
35InternVL2 4B· OpenVLM Leaderboard data68.6
36Claude 3 Haiku· OpenVLM Leaderboard data67.4
37Phi 4 Multimodal Instruct· OpenVLM Leaderboard data66.5
38Phi-3.5 Vision Instruct· OpenVLM Leaderboard data63.7
39Moondream2· OpenVLM Leaderboard data61.4
40kosmos-2· OpenVLM Leaderboard data54.7
41Molmo 7B-D· OpenVLM Leaderboard data33.6

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