MIOH - Counting: leaderboard
Metric: Accuracy (%; counting questions about how many objects there are across several images; multi-image questions built from COCO-ReM, PACO and Visual Genome scenes in three reasoning patterns (comprehensive, comparative, selective), averaged over the easy, hard-negative, hard-positive and eight-image conditions). Source: arxiv.org. Saturation forecast: Around September 2027. 29 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 57.5 |
| 2 | GPT-5 | 49.1 |
| 3 | MiniCPM-V-2.6 | 28.3 |
| 4 | Qwen 2.5 VL 7B | 28 |
| 5 | InternVL3.5-8B | 27.6 |
| 6 | Qwen 2 VL 7B | 26 |
| 7 | Phi-4 Multimodal Instruct | 24 |
| 8 | Qwen 2 VL 2B | 21.2 |
Interactive version: theaggregate.ai/benchmark?slug=mioh-counting · How It Works · Data refreshed daily, snapshot 2026-09-29.