Qwen3.8-27B on Perception Test
72.3%Rank #6 of 9 on Perception Test
- Coverage
- Full (454/454)
- Cost per task
- —
- Median latency
- 49.3s
- Tokens per task
- 2,574 in · 3,065 out
Run setting: 1 fps.
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 2,574 | 2,540 | 3,679 |
| of which cached | 0 | 0 | 0 |
| Output | 3,065 | 1,842 | 8,106 |
Over 454 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 49.3s
- Mean
- 71.8s
- 90th percentile
- 167.5s
- 99th percentile
- 314.9s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.