VideoZeroBench - Level-2 (Temporal 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-2 accuracy, with only the annotated time intervals given in text alongside the video, in percent; higher is better.

Metric: Level-2 accuracy (%), answering with only the annotated temporal intervals 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 A22B21.4#485
2Gemini 2.5 Pro21#276
3Gemini 3 Pro20.6#145
4Qwen 3.5 397B A17B20#264
5Seed 2.0 Pro18#222
6InternVL3.5-8B18#1604
7Qwen 3 VL 8B17.8#917
8Qwen 3 VL 4B16.6#1151
9GPT-5.214.6#185
10InternVL3.5-4B13.6#1930
11Video-R113.4#1379
12Video-o3-7B13.4
13Open-o3-Video-4B12.8
14Qwen 2.5 VL 7B12.4#1862
15VideoChat-R1.5-7B12.2#1254

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=videozerobench-level-2-temporal-evidence-given · How It Works · Data refreshed daily, snapshot 2026-11-21.