PushupBench: leaderboard
Metric: Exact match (%): share of clips whose predicted repetition count equals the ground truth, on PushupBench's 446 long-form exercise clips (22-117 s, 11 creators in 10 countries; 27 clips accept several valid counts) sampled at 5 fps up to 112 frames (Claude models capped at 100 frames by the API), 360p frames, greedy decoding, 10 prompt templates; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Flash | 42.1 |
| 2 | Gemini 3 Pro | 39.8 |
| 3 | Gemini 2.5 Pro | 29.9 |
| 4 | Qwen 3 VL 32B Instruct | 11.8 |
| 5 | GPT-5 | 10.9 |
| 6 | Gemini 2.5 Flash | 10.9 |
| 7 | Qwen 3 VL 32B (Thinking) | 9.9 |
| 8 | Qwen 3 VL 30B A3B (Thinking) | 9.6 |
| 9 | Claude Sonnet 4.5 | 9.5 |
| 10 | Qwen 2.5 VL 7B | 9.2 |
| 11 | Gemini 2.0 Flash | 8.9 |
| 12 | Qwen 3 VL 4B Instruct | 8.9 |
| 13 | Qwen 3 VL 235B A22B Instruct | 8.6 |
| 14 | Qwen 3 VL 4B (Thinking) | 8.2 |
| 15 | Qwen 3 VL 30B A3B Instruct | 7.6 |
Interactive version: theaggregate.ai/benchmark?slug=pushupbench · How It Works · Data refreshed daily, snapshot 2026-10-07.