Benchmarks
More
OverviewBenchmaxxingCompareChartsCost vs CapabilityProviders per ModelProvider explorerGatewaysEU & SovereignJev-class modelsImage JevBenchAbout
JevBench by Benchmark Heaven · v1.4.2.2 · individual system

Raw Qwen3 4B Instruct 2507 direct logits

Raw-logit control (base model) · by Alibaba Qwen / neutral reproduction

JevBench v1.4.2.2 score

41.0

Rank #22 of 91 ranked systems.

22.3 points behind Jev 1.13.0's 63.3.

Where it sits among the 91 ranked systems. The marked tick is Jev 1.13.0 (63.3).

Published axes

intelligence
46.4
calibration
29.1
speed
87.6
cost
59.7

Against Jev 1.13.0

Four radars compare this fixed pair across the score axes, accuracy per tier, and accuracy by family on the hard tier and sealed set. Further out is better on every spoke.

The four score axes

Radar: the four score axes, two systemsThe four score axes, Raw Qwen3 4B Instruct 2507 direct logits vs Jev 1.13.0. Intelligence: 46.4 vs 53.1; Calibration: 29.1 vs 76.3; Speed: 87.6 vs 83.3; Cost: 59.7 vs 52.0.50100Intelligence46.4 · 53.1Calibration29.1 · 76.3Speed87.6 · 83.3Cost59.7 · 52.0
0–100, the values in the table. A label-only system has no calibration (counted as 0).

Accuracy per tier, incl. sealed

Radar: accuracy per tier, incl. sealed, two systemsAccuracy per tier, incl. sealed, Raw Qwen3 4B Instruct 2507 direct logits vs Jev 1.13.0. Easy: 100% vs 100%; Standard: 86% vs 99%; Judge: 94% vs 95%; Hard: 52% vs 74%; Sealed: 27% vs 37%.50100Easy100% · 100%Standard86% · 99%Judge94% · 95%Hard52% · 74%Sealed27% · 37%
Share correct per tier; Sealed = the 308 private decisions, aggregate only.

Current question set by family (hard + sealed)

Radar: current question set by family (hard + sealed), two systemsCurrent question set by family (hard + sealed), Raw Qwen3 4B Instruct 2507 direct logits vs Jev 1.13.0. Ambiguous / abstain: 24% vs 43%; Judge: 46% vs 54%; Long policy: 27% vs 44%; Multi-hop: 40% vs 64%; Probability: 48% vs 63%; Temporal / numeric: 21% vs 28%; Trade-off: 34% vs 55%; Routing: 90% vs 100%; Trap / adversarial: 73% vs 83%; Paraphrase: 36% vs 64%; Safety judge: 31% vs 38%.50100Ambiguous /abstain24% · 43%Judge46% · 54%Long policy27% · 44%Multi-hop40% · 64%Probability48% · 63%Temporal /numeric21% · 28%Trade-off34% · 55%Routing90% · 100%Trap /adversarial73% · 83%Paraphrase36% · 64%Safety judge31% · 38%
Share correct per family across the 220 hard-tier decisions (public and held out) and the 308 sealed decisions of v1.4, pooled; Routing is hard-tier only, Paraphrase and Safety judge sealed only.

Sealed set by family

Radar: sealed set by family, two systemsSealed set by family, Raw Qwen3 4B Instruct 2507 direct logits vs Jev 1.13.0. Ambiguous / abstain: 24% vs 30%; Judge: 29% vs 34%; Long policy: 18% vs 28%; Multi-hop: 26% vs 45%; Paraphrase: 36% vs 64%; Probability: 39% vs 50%; Safety judge: 31% vs 38%; Temporal / numeric: 20% vs 29%; Trade-off: 35% vs 38%; Trap / adversarial: 42% vs 42%.50100Ambiguous /abstain24% · 30%Judge29% · 34%Long policy18% · 28%Multi-hop26% · 45%Paraphrase36% · 64%Probability39% · 50%Safety judge31% · 38%Temporal /numeric20% · 29%Trade-off35% · 38%Trap /adversarial42% · 42%
Share correct within each sealed family — system-level aggregates; the items stay private.

Availability and evidence

Openness
Code and weights marked open in the published row
License note
Apache-2.0
Cost evidence
estimated; the board’s row disclosure contains the published basis.
Endpoint condition
our lium.io RTX A6000 48 GB; local in-process; serial; one forward pass
Note on this row
Neutral raw-logit control.
Published source
https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507

From the public v1.4.2.2 aggregate. Scores and ranks can change when a new release is published.

Read the full board and published method. The overall score is a composite, not raw accuracy.