HoloCount - Action and State Attribution: 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 objects in a stated action or state (102 questions); higher is better). Source: arxiv.org. Saturation forecast: Around December 2026. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.5 397B A17B | 89.2 |
| 2 | Qwen 3.5 122B A10B | 89.2 |
| 3 | Qwen 3.5 27B | 87.3 |
| 4 | Gemini 3.1 Pro (Preview) | 85.3 |
| 5 | Qwen 3.5 4B | 85.3 |
| 6 | GPT-5.5 | 84.3 |
| 7 | Qwen 3.5 35B A3B | 84.3 |
| 8 | Qwen 3.5 9B | 83.3 |
| 9 | Gemini 3 Flash (Preview) | 82.4 |
| 10 | Qwen 3 VL 32B Instruct | 82.4 |
| 11 | Gemini 2.5 Pro | 81.4 |
| 12 | Kimi K2.5 | 80.4 |
| 13 | Kimi K2.6 | 80.4 |
| 14 | Qwen 3 VL 8B Instruct | 78.4 |
| 15 | Claude Opus 4.7 | 77.5 |
Interactive version: theaggregate.ai/benchmark?slug=holocount-action-and-state-attribution · How It Works · Data refreshed daily, snapshot 2026-09-29.