Image JevBench v0.3.0 · individual system
Visual-Jev 4B Answer-SFT
adapter · 4BBase model: Qwen/Qwen3-VL-4B-Instructsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
8.707
Rank #25 of 44 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 24.0
- calibration
- 47.7
- speed
- 84.2
- cost
- 47.4
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.056510
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.149 s / 0.349 s
- Adjusted latency · p50 / p95
- 0.448 s / 0.848 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 49.8%
- Sealed accuracy
- 48.2%
Published 2026-10-06. Scores and ranks can change in a later release. See the method notes.