HoloCount - Joint-Set Aggregation: leaderboard
Giving the total count of objects from several categories: a subset of HoloCount (Meituan, 2026), a visual counting benchmark for multimodal LLMs. Models answer with a single integer, zero-shot; exact-match accuracy in percent, higher is better.
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; the total count of objects from several categories (100 questions); higher is better). Source: arxiv.org. Saturation forecast: Not forecast. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 82 | #94 |
| 2 | Qwen 3.5 122B A10B | 82 | #331 |
| 3 | Qwen 3.5 27B | 79 | #359 |
| 4 | Qwen 3.5 397B A17B | 78 | #246 |
| 5 | Qwen 3.5 35B A3B | 77 | #482 |
| 6 | Gemini 3 Flash (Preview) | 77 | #144 |
| 7 | Qwen 3.5 4B | 71 | #1175 |
| 8 | Qwen 3.5 9B | 69 | #803 |
| 9 | Kimi K2.6 | 68 | #178 |
| 10 | Kimi K2.5 | 67 | #256 |
| 11 | Gemini 2.5 Pro | 65 | #266 |
| 12 | Claude Opus 4.7 | 62 | #79 |
| 13 | Qwen 2.5 VL 72B Instruct | 61 | #846 |
| 14 | GPT-5.5 | 60 | #49 |
| 15 | Claude Opus 4.8 | 58 | #56 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=holocount-joint-set-aggregation · How It Works · Data refreshed daily, snapshot 2026-11-13.