vjev-vision (yah01)
adapter · 4BBase model: Qwen/Qwen3.5-4Bsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
8.857
Rank #24 of 44 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 21.7
- calibration
- 89.6
- speed
- 89.7
- cost
- 59.0
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.023266
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.068 s / 0.111 s
- Adjusted latency · p50 / p95
- 0.286 s / 0.371 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 45.5%
- Sealed accuracy
- 48.1%
Published 2026-10-06. Scores and ranks can change in a later release. See the method notes.