HoloCount - Atomic Counting: 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; atomic counting of one object category (182 questions); higher is better). Source: arxiv.org. Saturation forecast: Around December 2026. 30 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)95.1
2Gemini 3 Flash (Preview)93.4
3Qwen 3.5 397B A17B92.9
4Qwen 3.5 35B A3B90.7
5Qwen 3.5 27B90.7
6Qwen 3.5 122B A10B90.7
7Kimi K2.589
8Kimi K2.689
9Qwen 3.5 9B86.8
10GPT-5.584.6
11Gemini 2.5 Pro84.1
12Qwen 3 VL 32B Instruct84.1
13Claude Opus 4.882.4
14Qwen 3.5 4B80.8
15Qwen 2.5 VL 72B Instruct80.8

Interactive version: theaggregate.ai/benchmark?slug=holocount-atomic-counting · How It Works · Data refreshed daily, snapshot 2026-09-29.