HR-Bench (8K)
Multimodal · 2024-08-28
HR-Bench tests fine-grained perception in 8K-resolution images with 200 four-option questions: 100 fine-grained single-instance (FSP) and 100 cross-instance (FCP) perception items. Each question is asked under four cyclic orderings of its options and the score is accuracy averaged over the orderings.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-4B | 82.1 |
| LFM3-VL-2B | 76.2 |
| Gemini 3.5 Flash-Lite | 75.7 |
| Qwen3.5-2B | 74.4 |
| Gemma 4 12B IT | 70.6 |
| GPT-5.6 Luna | 60.0 |
| Gemma 4 E2B IT | 57.0 |