Image JevBench v0.3.0 · individual system
NeoHorse Jev 4B
4BBase model: NeoHorse-1-4Bsourcelocal/GPU evaluation
Image JevBench v0.3.0 composite score
30.452
Rank #10 of 55 ranked systems.
From the published full-benchmark aggregate. The composite combines intelligence, calibration, speed and cost.
Published axes
- intelligence
- 36.8
- calibration
- 78.4
- speed
- 86.8
- cost
- 50.4
Cost and speed evidence
- Cost per 1,000 decisions
- USD 0.041947
- Cost basis
- measured GPU seconds x $1.19/GPU-hour
- Measured latency · p50 / p95
- 0.140 s / 0.169 s
- Adjusted latency · p50 / p95
- 0.429 s / 0.488 s
- Latency adjustment
- v1.4 local latency adjustment
- Public accuracy
- 57.0%
- Sealed accuracy
- 57.6%
- Measurement note
- Official ImageJev v0.3 S1200+P300 pass 1 on 7 Oct 2026 on one RTX PRO 6000 with the same reviewed runtime as the other self-hosted rows. The reviewed Hugging Face revision 56c36ae6 no longer exists (the author rewrote the history), so the run used revision 434cb21d. Its four weight files, eight package source files and wheel are byte-identical to the reviewed ones; only README.md and SHA256SUMS differ. 1,480 ok, 20 runtime errors counted wrong.
Published 2026-10-12. Scores and ranks can change in a later release. See the method notes.