A3Bench - Basic Questions: leaderboard

Zero-shot multiple-choice video QA accuracy (%) on single-attribute basic questions in A3Bench (arXiv:2611.24762).

Metric: Basic Questions Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2028. 14 models tracked. Updated 20 Nov 2026.

Top models

#ModelScoreOverall rank
1InternVL2.5-8B52.61#2370
2Qwen 3 VL 8B Instruct51.07#1035
3Ovis2.5-9B50.62#1189
4InternVL3-9B50.05#1731
5VideoLLaMA3-7B49.65#2000
6Gemma 4 26B A4B (IT)49.07#561
7InternVL3.5-8B48.34#1599
8InternVideo2.5-8B48.07
9GPT-5.145.63#242
10Qwen 3 VL 8B (Thinking)44.59#920
11Qwen 2.5 VL 7B Instruct41.87#1904
12ViCLIP28.11
13X-CLIP23.67
14LLaVA-Video-7B-Qwen219.58#1977

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

Interactive version: theaggregate.ai/benchmark?slug=a3bench-basic-questions · How It Works · Data refreshed daily, snapshot 2026-11-20.