VirtueBench (1024 Frames): leaderboard
Metric: Accuracy (%) over VirtueBench's 587 open-ended long-video questions answered from 1024 uniformly sampled frames (at most 512x512 pixels), each with a reference answer for that frame level: an answerable question counts when a GPT-4o judge finds the answer consistent with the reference, an unanswerable one only when the model says the frames lack the information; the prompt tells the model not to guess; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 53.66 | #237 |
| 2 | Qwen 3 VL 32B Instruct | 32.54 | #276 |
| 3 | Qwen 2.5 VL 72B Instruct | 31.52 | #364 |
| 4 | Qwen 2.5 VL 32B Instruct | 31.52 | #443 |
| 5 | Qwen 3 VL 8B Instruct | 29.98 | #401 |
| 6 | Qwen 3 VL 32B (Thinking) | 25.89 | #287 (Qwen 3 VL 32B) |
| 7 | Qwen 2.5 VL 7B Instruct | 25.55 | #643 |
| 8 | Qwen 3 VL 8B (Thinking) | 23.34 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=virtuebench-1024-frames · How It Works · Data refreshed daily, snapshot 2026-10-11.