Qwen3.8-27B on NExT-QA
82.2%Rank #8 of 9 on NExT-QA
- Coverage
- Full (4996/4996)
- Cost per task
- —
- Median latency
- 8.1s
- Tokens per task
- 4,504 in · 791 out
Run setting: 1 fps.
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 4,504 | 3,734 | 8,955 |
| of which cached | 0 | 0 | 0 |
| Output | 791 | 306 | 1,947 |
Over 5,006 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 8.1s
- Mean
- 17.7s
- 90th percentile
- 39.0s
- 99th percentile
- 131.6s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.