Qwen3.5-35B-A3B on NExT-QA
83.4%Rank #7 of 9 on NExT-QA
- Coverage
- Full (4996/4996)
- Cost per task
- —
- Median latency
- 5.3s
- Tokens per task
- 4,466 in · 1,270 out
Run setting: 1 fps.
Tokens and latency
Tokens per call
| Token type | Mean | Median | 90th pct. |
|---|---|---|---|
| Input | 4,466 | 3,696 | 8,917 |
| of which cached | 0 | 0 | 0 |
| Output | 1,270 | 589 | 2,808 |
Over 5,006 calls that returned an answer, including questions asked again after a failed attempt.
Latency
- Median
- 5.3s
- Mean
- 9.0s
- 90th percentile
- 18.9s
- 99th percentile
- 50.4s
We only consider the first question against a video for self hosted models, ignoring prefix caching latency.