HoloCount - Partial Object Occlusion: leaderboard
Metric: Accuracy (%; exact-match accuracy of the single integer count a multimodal model returns zero-shot for an image and a counting question, with a strict integer-only system prompt; counting heavily occluded objects (103 questions); higher is better). Source: arxiv.org. Saturation forecast: Around 2029. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.5 35B A3B | 65 |
| 2 | Qwen 3.5 9B | 64.1 |
| 3 | Qwen 3.5 122B A10B | 64.1 |
| 4 | Gemini 3.1 Pro (Preview) | 63.1 |
| 5 | Kimi K2.6 | 61.2 |
| 6 | Qwen 3.5 397B A17B | 61.2 |
| 7 | Qwen 3.5 27B | 61.2 |
| 8 | Gemini 3 Flash (Preview) | 61.2 |
| 9 | Qwen 2.5 VL 72B Instruct | 58.3 |
| 10 | Gemini 2.5 Pro | 56.3 |
| 11 | Claude Opus 4.8 | 55.3 |
| 12 | Qwen 3.5 4B | 55.3 |
| 13 | GPT-4.1 | 50.5 |
| 14 | Claude Opus 4.7 | 50.5 |
| 15 | Qwen 3 VL 32B Instruct | 47.6 |
Interactive version: theaggregate.ai/benchmark?slug=holocount-partial-object-occlusion · How It Works · Data refreshed daily, snapshot 2026-09-29.