Qwen3.8-27B on Video-MME v2
50.5%Rank #6 of 9 on Video-MME v2
- Coverage
- Full (3200/3200)
- Cost per task
- —
- Median latency
- 86.5s
- Tokens per task
- 56,163 in · 4,653 out
Run setting: 1 fps, 768-frame cap (226 videos thinned).
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 56,163 | 56,201 | 77,051 |
| of which cached | 0 | 0 | 0 |
| Output | 4,653 | 3,661 | 10,067 |
Over 3,468 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 86.5s
- Mean
- 113.7s
- 90th percentile
- 231.7s
- 99th percentile
- 404.6s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.