OmniClean: leaderboard
OmniClean (StepFun-Audio Team, 2026) is a visually debiased view of nine audio-visual-language benchmarks: a query is removed when Qwen3-VL-30B-A3B-Thinking answers it correctly in any of 16 attempts from the video or image and question alone, with the audio withheld. Models answer under each source benchmark's official protocol; the score is the unweighted mean accuracy over the nine benchmarks (seven cleaned, AV-Odyssey and a CG-AV-Counting subset kept whole) in percent, higher is better.
Metric: Accuracy (%; unweighted mean over nine audio-visual benchmarks, seven reduced to the queries a visual-only probe could not solve and two kept whole, 8,551 queries; official answer matching). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3 Omni 30B A3B Instruct | 30.51 | #902 |
| 2 | Qwen2.5-Omni-7B | 27.39 | #1797 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omniclean · How It Works · Data refreshed daily, snapshot 2026-11-12.