TempCloze: leaderboard
The middle of a video clip is cut out and the model has to say what belonged there — what happened, who did it, and in which order. Language shortcuts are the thing it is built to remove, so the options cannot be ranked by wording. Humans score 97 on the dimension mean; most open models sit near the 25 the random baseline gets.
Metric: Mean Accuracy (%). Source: arxiv.org. Saturation forecast: Around January 2027. 32 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 Plus (Thinking) | 85.74 | #125 (Qwen 3.5 Plus) |
| 2 | Qwen 3.5 397B A17B (Thinking) | 68.27 | #146 (Qwen 3.5 397B A17B) |
| 3 | Qwen 3.5 35B A3B (Thinking) | 65.54 | #256 (Qwen 3.5 35B A3B) |
| 4 | Kimi K2.5 (Non-reasoning) | 59.41 | #144 (Kimi K2.5) |
| 5 | Qwen 3 VL 32B Instruct | 41.84 | #283 |
| 6 | Qwen 3 VL 8B Instruct | 30.55 | #420 |
| 7 | Qwen 2.5 VL 7B Instruct | 25.87 | #654 |
| 8 | Qwen 3 VL 32B (Thinking) | 25.38 | #294 (Qwen 3 VL 32B) |
| 9 | Qwen 3 VL 4B (Thinking) | 25.2 | #480 (Qwen 3 VL 4B) |
| 10 | Qwen 3 VL 8B (Thinking) | 24.87 | |
| 11 | Qwen 3 VL 4B Instruct | 24.19 | #515 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=tempcloze · How It Works · Data refreshed daily, snapshot 2026-10-20.