VideoZeroBench - Level-1 (Temporal and Spatial Evidence Given): leaderboard

VideoZeroBench (2026) asks 500 manually annotated, deliberately hard open-ended questions with exact short answers about 138 long videos in 13 domains, each with annotated temporal intervals and bounding boxes as supporting evidence. Level-1 accuracy, with the annotated time intervals and key spatial regions given in text alongside the video, in percent; higher is better.

Metric: Level-1 accuracy (%), answering with the annotated temporal intervals and key spatial regions given in text, on VideoZeroBench's 500 manually annotated open-ended questions (exact short answers) over 138 long videos in 13 domains; Gemini models get the raw video through the API, the others 1 fps frames up to the printed cap, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 235B A22B28.4#486
2Seed 2.0 Pro25.4#226
3Gemini 2.5 Pro25.2#280
4Qwen 3.5 397B A17B25#270
5Qwen 3 VL 8B24.8#927
6Gemini 3 Pro23.6#146
7Qwen 3 VL 4B22.2#1163
8InternVL3.5-8B19.4#1618
9Open-o3-Video-4B17.6
10GPT-5.217.4#190
11VideoChat-R1.5-7B17.4#1260
12Video-R117.2#1384
13Gemini 2.5 Flash17#466
14Qwen 2.5 VL 7B16.2#1878
15Video-o3-7B15

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=videozerobench-level-1-temporal-and-spatial-evidence-given · How It Works · Data refreshed daily, snapshot 2026-11-22.