Qwen3.8-27B on LVBench
64.6%Rank #5 of 9 on LVBench
- Coverage
- Full (1549/1549)
- Cost per task
- —
- Median latency
- 43.0s
- Tokens per task
- 76,990 in · 2,918 out
Run setting: 1 fps, 768-frame cap (103 videos thinned).
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 76,990 | 76,859 | 76,935 |
| of which cached | 0 | 0 | 0 |
| Output | 2,918 | 1,570 | 7,527 |
Over 1,549 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 43.0s
- Mean
- 67.1s
- 90th percentile
- 132.8s
- 99th percentile
- 282.0s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.