VSTAT - Count: leaderboard

Metric: Score (%) on questions about count-type state elements: 1,500 questions on 834 synthetic (Blender) and real-world video clips that require tracking a visual state through the whole video; accuracy on multiple-choice questions and mean relative accuracy on numerical questions (%), open models at their best of 16 to 128 uniformly sampled frames; higher is better. Source: arxiv.org. Saturation forecast: Around 2032. 23 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview) (Low)42.6
2Gemini 3.1 Pro (Preview) (High)42.1
3Gemini 3 Flash (High)33.2
4Qwen 3 VL 8B Instruct30.9
5Molmo2-8B30.9
6Qwen 3 VL 4B Instruct27
7Qwen 3 VL 8B (Thinking)25.9
8InternVL3.5-8B25.1
9GLM-4.1V-9B (Thinking)24.8
10Qwen 3 VL 4B (Thinking)21

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