Glance (frozen Qwen3-VL-4B)
4BBase model: Qwen3-VL-4Bsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
6.854
Rank #28 of 44 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 21.3
- calibration
- 35.6
- speed
- 87.8
- cost
- 52.8
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.035020
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.103 s / 0.160 s
- Adjusted latency · p50 / p95
- 0.356 s / 0.469 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 46.8%
- Sealed accuracy
- 47.0%
Published 2026-10-06. Scores and ranks can change in a later release. See the method notes.