K9-Bench - Interaction Analysis: leaderboard
Metric: MCQ accuracy (%; the Interaction Analysis category of K9-Bench, 4,744 five-option questions on 907 real-world dog videos; the free-form CoT answer counts as correct when its Qwen3-Embedding-8B similarity to the gold option exceeds 0.5 and no wrong option is closer; 32 uniformly sampled frames for GPT-4o and the open models, the full video at 1 FPS for Gemini 2.5 Pro and Qwen3-VL-235B-A22B). Source: arxiv.org. Saturation forecast: Around August 2028. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Pro | 42.3 |
| 2 | Qwen 3 VL 235B A22B Instruct | 39.4 |
| 3 | InternVL3.5-8B | 37.5 |
| 4 | Qwen 3 VL 32B Instruct | 35 |
| 5 | GPT-4o (2024-08-06) | 33.9 |
| 6 | Qwen 3 VL 32B (Thinking) | 32.1 |
| 7 | Qwen 3 VL 8B Instruct | 31.9 |
| 8 | Qwen 3 VL 4B Instruct | 30.9 |
| 9 | Qwen 3 VL 8B (Thinking) | 29.3 |
| 10 | Qwen 3 VL 4B (Thinking) | 26.2 |
Interactive version: theaggregate.ai/benchmark?slug=k9-bench-interaction-analysis · How It Works · Data refreshed daily, snapshot 2026-09-29.