VSTAT - Location: leaderboard

Metric: Score (%) on questions about location-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 2035. 23 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview) (High)41.6
2Gemini 3.1 Pro (Preview) (Low)38.5
3Qwen 3 VL 8B Instruct37
4Molmo2-8B37
5Gemini 3 Flash (High)36.6
6GLM-4.1V-9B (Thinking)33.9
7Qwen 3 VL 4B Instruct33.3
8InternVL3.5-8B33.2
9Qwen 3 VL 8B (Thinking)32.2
10Qwen 3 VL 4B (Thinking)31.8

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