V* Bench
Multimodal · 2023-12-21
V* Bench evaluates visual search in high-resolution images with 191 multiple-choice questions on small-object attribute recognition and relative spatial position. This row holds no-tool runs; runs with a code interpreter for cropping and zooming use the with-Python row.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.6 27B | 94.7 |
| Qwen3.6 Plus (2026-04-02) | 90.5 |
| Qwen3.5 122B A10B | 90.1 |
| Qwen3.6 35B A3B | 90.1 |
| Qwen3.5-4B | 89.5 |
| Qwen3.5 35B A3B | 89.5 |
| Qwen3.5-9B | 88.5 |
| Gemini 3 Pro Preview | 88.0 |
| LFM3-VL-2B | 86.4 |
| Qwen3.5-2B | 85.9 |
| GPT-5.6 Luna | 79.6 |
| Kimi K2.5 | 77.0 |
| GPT-5.2 | 75.9 |
| Gemini 3.5 Flash-Lite | 74.9 |
| GPT-4o (source unspecified) | 73.9 |
| GPT-5 Mini | 71.7 |
| o1 (2024-12-17) | 69.7 |
| Gemma 4 12B IT | 69.6 |
| Gemini 2.5 Flash-Lite | 69.6 |
| GPT-5 Nano | 68.1 |
| Opus 4.5 | 67.0 |
| Sonnet 4.5 | 58.6 |
| Gemma 4 E2B IT | 44.5 |