TraceAV-Bench (Visual Only): leaderboard
TraceAV-Bench (2026), long audio-visual video questions with multi-hop evidence, answered from the video frames alone: vision-only models and omni models with the audio stream removed. The score is accuracy in percent, the macro-average of 12 general sub-tasks; higher is better.
Metric: Accuracy (%), macro-average of the 12 general sub-tasks, four-option multiple choice over long audio-visual videos (10 to 140 minutes), exact match of the selected option set, each model at its official inference setting, with visual input only: video frames without the audio track (vision-only MLLMs, and omni models with the audio stream removed); higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B | 41.04 | #622 |
| 2 | Ming-Flash-Omni-2.0 | 38.55 | #1266 |
| 3 | Qwen 3 Omni 30B A3B | 37.4 | #847 |
| 4 | Qwen 3 VL 8B | 34.31 | #917 |
Interactive version: theaggregate.ai/benchmark?slug=traceav-bench-visual-only · How It Works · Data refreshed daily, snapshot 2026-11-21.