HR-Bench (4K)
Multimodal · 2024-08-28
The 4K version of HR-Bench, built from the same 200 8K images and asking the same 100 single-instance (FSP) and 100 cross-instance (FCP) four-option questions at 4K resolution. Each question is asked under four cyclic orderings of its options and the score is accuracy averaged over the orderings; it shares the 8K family because the questions are the same.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.5-4B | 88.1 |
| LFM3-VL-2B | 82.3 |
| GPT-5.6 Luna | 80.1 |
| Gemini 3.5 Flash-Lite | 79.2 |
| Gemma 4 12B IT | 77.4 |
| Qwen3.5-2B | 76.9 |
| Gemma 4 E2B IT | 65.5 |