Benchmark
HallusionBench
Vision benchmarks measure how well models understand images. Scores are percentages. Higher is better.
- Models measured
- 64
- Scores recorded
- 66
- Sources
- 3
- IQ weight
- 0.90
Score distribution
- 53.3%
- Mean
- 53.9%
- Median
- 10.7%
- Std dev
- 40.5%
- 10th percentile
- 67.1%
- 90th percentile
- 19.8%
- Min
This benchmark has a weight of 0.90 for chat models, measured over 64 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 51 | DeepSeek VL2· OpenVLM Leaderboard data | 45.3% |
| 52 | GLM 4 Vision· OpenVLM Leaderboard data | 44.9% |
| 53 | Llama 3.2 90B Vision Instruct· OpenVLM Leaderboard data | 44.1% |
| 54 | InternVL2 4B· OpenVLM Leaderboard data | 42.4% |
| 55 | Claude 3 Sonnet· OpenVLM Leaderboard data | 41.3% |
| 56 | Qwen-VL Max· OpenVLM Leaderboard data | 41.2% |
| 57 | Gemma 3 4B· OpenVLM Leaderboard data | 40.8% |
| 58 | Qwen-VL Plus· OpenVLM Leaderboard data | 40.6% |
| 59 | Phi 4 Multimodal Instruct· OpenVLM Leaderboard data | 40.5% |
| 60 | Phi-3.5 Vision Instruct· OpenVLM Leaderboard data | 40.5% |
| 61 | Llama 3.2 11B Vision Instruct· OpenVLM Leaderboard data | 40.3% |
| 62 | Claude 3 Haiku· OpenVLM Leaderboard data | 39.2% |
| 63 | Claude 3 Opus· OpenVLM Leaderboard data | 37.8% |
| 64 | GPT-4.1 nano· OpenVLM Leaderboard data | 36.9% |
| 65 | Moondream2· OpenVLM Leaderboard data | 33% |
| 66 | kosmos-2· OpenVLM Leaderboard data | 19.8% |
Showing 51–66 of 66 results.
Data credits: OpenVLM Leaderboard (OpenCompass / VLMEvalKit), Apache-2.0 (licence) · Data by LLM Stats · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.