Image JevBench v0.3.0 · individual system
H2O-Lightning-4B v1.4.1 @d7f1b05a (image)
LLM decoderBase model: Qwen/Qwen3.5-4Bsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
68.514
Listed, not ranked.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 55.5
- calibration
- 88.2
- speed
- 88.7
- cost
- 56.3
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.028525
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.086 s / 0.133 s
- Adjusted latency · p50 / p95
- 0.323 s / 0.416 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 67.2%
- Sealed accuracy
- 71.5%
- Measurement note
- same weights (c9b6290d); serving settings changed after the 27 Oct ImageJev result was published (image temperature 0.65 -> 1.0, BM25 selector, transformers 5.20.2); second read of the same frozen input
Published 2026-10-28. Scores and ranks can change in a later release. See the method notes.