HoloCount - Complementary Exclusion: 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 one set but not another (101 questions); higher is better). Source: arxiv.org. Saturation forecast: Around January 2027. 30 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 84.2 |
| 2 | Qwen 3.5 397B A17B | 80.2 |
| 3 | Gemini 3 Flash (Preview) | 80.2 |
| 4 | Qwen 3.5 122B A10B | 79.2 |
| 5 | Qwen 3.5 35B A3B | 78.2 |
| 6 | Qwen 3.5 27B | 76.2 |
| 7 | Gemini 2.5 Pro | 74.3 |
| 8 | Kimi K2.6 | 74.3 |
| 9 | Qwen 3.5 4B | 70.3 |
| 10 | Qwen 3.5 9B | 67.3 |
| 11 | GPT-5.5 | 66.3 |
| 12 | Kimi K2.5 | 63.4 |
| 13 | Claude Opus 4.8 | 63.4 |
| 14 | GPT-5.4 | 61.4 |
| 15 | GPT-4.1 | 60.4 |
Interactive version: theaggregate.ai/benchmark?slug=holocount-complementary-exclusion · How It Works · Data refreshed daily, snapshot 2026-09-29.