Qwen3.8-27B on EgoSchema
79.1%Rank #5 of 9 on EgoSchema
- Coverage
- Full (500/500)
- Cost per task
- —
- Median latency
- 23.1s
- Tokens per task
- 18,346 in · 1,390 out
Run setting: 1 fps.
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 18,346 | 18,321 | 18,490 |
| of which cached | 0 | 0 | 0 |
| Output | 1,390 | 778 | 3,453 |
Over 500 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 23.1s
- Mean
- 35.9s
- 90th percentile
- 78.9s
- 99th percentile
- 172.1s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.