Image JevBench v0.3.0 · individual system
JPT-4B (kirp / llm2jev)
adapter · 4BBase model: Qwen3.5-4Bsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
26.549
Rank #8 of 44 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 38.3
- calibration
- 73.0
- speed
- 83.3
- cost
- 45.7
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.064571
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.176 s / 0.388 s
- Adjusted latency · p50 / p95
- 0.502 s / 0.926 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 57.0%
- Sealed accuracy
- 59.1%
Published 2026-10-06. Scores and ranks can change in a later release. See the method notes.