JevBench is a benchmark maintained by Benchmark Heaven for “Jev-class” typed decision models. Its repository explicitly says it is not affiliated with or endorsed by TypeSafe AI, whose Jev model is one of the systems tested.
The current v1.4.2 board ranks decider-4b v2 first at 64.13 and Jev 1.13.0 second at 63.29. The scoring method combines 20% chance-corrected performance on 308 new sealed decisions with 80% of the earlier Intelligence score, blends calibration, uses an equal-weight harmonic mean across four axes, and penalizes large public-to-sealed accuracy gaps.
Sealing some items addresses a real weakness of public benchmarks: developers can tune against visible questions. But the hidden set is still only 308 decisions, and the benchmark owner says only aggregates are published. This makes the measurements harder to independently reproduce. The score also reflects the benchmark’s chosen tasks and weights; JevBench itself recommends looking at the Intelligence axis when raw reasoning is the question.
The version change matters: an earlier draft of this story called Jev the leader, which is no longer true on v1.4.2. The result is a useful snapshot under one protocol, not a universal ordering of decision models.