BLINK
Multimodal · 2024-04-18
BLINK reformats 14 classic computer-vision perception tasks (relative depth, correspondence, forensics, multi-view reasoning and others) into 3,807 multiple-choice questions that humans answer at a glance but multimodal models find hard. Scores report accuracy, normally on the 1,901-question validation split.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K2.5 | 78.9 |
| Gemini 3 Pro Preview | 78.7 |
| LFM3-VL-2B | 71.1 |
| GPT-5 | 71.0 |
| Gemma 4 12B IT | 70.9 |
| Qwen3-VL-235B-A22B-Instruct | 70.7 |
| Gemini 2.5 Pro | 70.6 |
| Qwen3.5-4B | 69.8 |
| Qwen3-VL-8B-Instruct | 69.1 |
| Gemini 3.5 Flash-Lite | 68.9 |
| Opus 4.5 | 68.8 |
| GPT-5.6 Luna | 68.6 |
| Qwen3 VL 32B Thinking | 68.5 |
| Gemini 2.5 Flash | 68.1 |
| Qwen3 VL 30B A3B Instruct | 67.7 |
| Qwen3 VL 32B Instruct | 67.3 |
| Qwen3 VL 4B Instruct | 65.8 |
| Qwen3 VL 30B A3B Thinking | 65.4 |
| GLM-4.5V | 65.3 |
| Qwen3 VL 8B Thinking | 64.7 |
| Qwen2.5-VL-72B-Instruct | 64.4 |
| Opus 4.1 | 64.1 |
| Qwen3.5-2B | 63.5 |
| Qwen3 VL 4B Thinking | 63.4 |
| LFM2.5-VL-3B | 61.5 |
| Sonnet 4 | 60.4 |
| GPT-5 Nano | 58.3 |
| Nemotron Nano 12B V2 VL BF16 | 57.6 |
| Kimi-VL-A3B-Instruct | 57.3 |
| Qwen3 VL 2B Thinking | 57.2 |
| GPT-5 Mini | 56.7 |
| Qwen2.5-VL-7B-Instruct | 56.4 |
| Qwen3 VL 2B Instruct | 53.8 |
| GPT-4o Mini | 53.6 |
| InternVL3-2B | 53.1 |
| Gemma 4 E2B IT | 52.9 |
| LFM2-VL-3B | 51.0 |
| Gemma 3 12B IT | 50.3 |
| Qwen2.5-VL-3B-Instruct | 49.0 |
| LFM2-VL-1.6B | 44.5 |