M3-AD-Bench: leaderboard
Metric: Balanced accuracy (%) of binary anomaly detection (is the image defective?), the mean of the recall on normal and on anomalous images, averaged over the Texture, Workpiece and Electronic scenes, zero-shot on M3-AD-Bench, the evaluation split of M3-AD (industrial images re-annotated from public anomaly-detection datasets under one taxonomy; its categories are held out from the M3-AD-FT training split), one structured answer per image; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 32B Instruct | 79.5 | #276 |
| 2 | Qwen 3 VL 4B Instruct | 72.1 | #506 |
| 3 | Qwen 2.5 VL 72B Instruct | 70.7 | #364 |
| 4 | Gemini 2.5 Flash Lite | 65.8 | #413 |
| 5 | Qwen 3 VL 8B Instruct | 63.7 | #401 |
| 6 | Qwen 3 VL 8B (Thinking) | 62.9 | |
| 7 | Qwen 2.5 VL 7B Instruct | 62.3 | #643 |
| 8 | Qwen 2 VL 7B Instruct | 59 | #816 |
| 9 | Qwen 3 VL 4B (Thinking) | 57.3 | #471 (Qwen 3 VL 4B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=m3-ad-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.