OpenVLM BLINK: leaderboard
BLINK overall accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice visual perception tasks such as relative depth, correspondence, counting and forensics that humans solve at a glance. Accuracy in percent, higher is better; exact-zero rows (failed runs) are dropped.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 182 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 68.5 | #257 |
| 2 | InternVL3-78B | 66.4 | #361 |
| 3 | InternVL3-38B | 64.2 | #413 |
| 4 | InternVL2.5-78B | 63.8 | #434 |
| 5 | Qwen 2 VL 72B | 60.5 | #450 |
| 6 | InternVL3-14B | 59.9 | #506 |
| 7 | Gemini 1.5 Pro | 59.1 | #454 |
| 8 | GPT-4.1 Mini | 59 | #363 |
| 9 | Gemini 1.5 Flash | 57.7 | #620 |
| 10 | Claude 3.7 Sonnet | 56.6 | #259 |
| 11 | Moonshot v1 8K | 55.8 | #622 |
| 12 | InternVL3-8B | 55.4 | #624 |
| 13 | Qwen 2.5 VL 7B | 55.3 | #641 |
| 14 | MiniCPM-V-2.6 | 55.2 | #827 |
| 15 | Qwen 2 VL 7B | 53.8 | #743 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-blink · How It Works · Data refreshed daily, snapshot 2026-11-05.