HoloCount: leaderboard
HoloCount (Meituan, 2026) is a visual counting benchmark for multimodal LLMs with 2,480 questions in 20 subsets that span semantic counting, analytical counting and robustness tests. Models answer with a single integer, zero-shot; exact-match accuracy in percent over all questions, higher is better.
Metric: Overall 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, over all 2,480 questions in the 20 subsets; higher is better). Source: arxiv.org. Saturation forecast: Not forecast. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 76.9 | #246 |
| 2 | Qwen 3.5 27B | 75.9 | #359 |
| 3 | Qwen 3.5 122B A10B | 75.4 | #331 |
| 4 | Qwen 3.5 35B A3B | 75 | #482 |
| 5 | Gemini 3 Flash (Preview) | 74.8 | #144 |
| 6 | Gemini 3.1 Pro (Preview) | 74.7 | #94 |
| 7 | Kimi K2.6 | 73 | #178 |
| 8 | Qwen 3.5 9B | 71.4 | #803 |
| 9 | Kimi K2.5 | 68.8 | #256 |
| 10 | Qwen 3.5 4B | 68.6 | #1175 |
| 11 | Gemini 2.5 Pro | 67.9 | #266 |
| 12 | GPT-5.5 | 67.6 | #49 |
| 13 | Claude Opus 4.7 | 62.8 | #79 |
| 14 | Qwen 3 VL 32B Instruct | 62.7 | #585 |
| 15 | Claude Opus 4.8 | 62.6 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-13.