Qwen3.8-27B on Q-Bench Video
67.9%Rank #5 of 9 on Q-Bench Video
- Coverage
- Full (868/868)
- Cost per task
- —
- Median latency
- 7.8s
- Tokens per task
- 1,139 in · 780 out
Run setting: 1 fps.
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 1,139 | 1,093 | 2,155 |
| of which cached | 0 | 0 | 0 |
| Output | 780 | 366 | 1,738 |
Over 859 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 7.8s
- Mean
- 16.3s
- 90th percentile
- 35.8s
- 99th percentile
- 145.1s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.