HoloCount: leaderboard
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: Around January 2028. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.5 397B A17B | 76.9 |
| 2 | Qwen 3.5 27B | 75.9 |
| 3 | Qwen 3.5 122B A10B | 75.4 |
| 4 | Qwen 3.5 35B A3B | 75 |
| 5 | Gemini 3 Flash (Preview) | 74.8 |
| 6 | Gemini 3.1 Pro (Preview) | 74.7 |
| 7 | Kimi K2.6 | 73 |
| 8 | Qwen 3.5 9B | 71.4 |
| 9 | Kimi K2.5 | 68.8 |
| 10 | Qwen 3.5 4B | 68.6 |
| 11 | Gemini 2.5 Pro | 67.9 |
| 12 | GPT-5.5 | 67.6 |
| 13 | Claude Opus 4.7 | 62.8 |
| 14 | Qwen 3 VL 32B Instruct | 62.7 |
| 15 | Claude Opus 4.8 | 62.6 |
Interactive version: theaggregate.ai/benchmark?slug=holocount · How It Works · Data refreshed daily, snapshot 2026-09-29.