LLM Stats (CharadesSTA): leaderboard
Temporal activity grounding in videos — given a sentence query, models must localize the start and end times of the described activity in a video.
Metric: Score (%). Source: llm-stats.com. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B Instruct | 64.8 | #273 |
| 2 | Qwen 3 VL 30B A3B Instruct | 63.5 | #384 |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 63.5 | #238 (Qwen 3 VL 235B A22B) |
| 4 | Qwen 3 VL 32B (Thinking) | 62.8 | #294 (Qwen 3 VL 32B) |
| 5 | Qwen 3 VL 30B A3B (Thinking) | 62.7 | #351 (Qwen 3 VL 30B A3B) |
| 6 | Qwen 3 VL 32B Instruct | 61.2 | #283 |
| 7 | Qwen 3 VL 8B (Thinking) | 59.9 | |
| 8 | Qwen 3 VL 4B (Thinking) | 59 | #480 (Qwen 3 VL 4B) |
| 9 | Qwen 3 VL 8B Instruct | 56 | #420 |
| 10 | Qwen 3 VL 4B Instruct | 55.5 | #515 |
| 11 | Qwen 2.5 VL 32B Instruct | 54.2 | #453 |
| 12 | Qwen 2.5 VL 7B Instruct | 43.6 | #654 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=llm-stats-charadessta · How It Works · Data refreshed daily, snapshot 2026-10-20.