# JevBench

As of Oct 14, 2026, JevBench (v1.6.1) is maintained by Benchmark Heaven. Metric capability or headline composite. Ceiling 100. Saturation unknown. Contamination risk med.

Declared cadence: revisioned. Catalog snapshots: 1, latest Oct 8, 2026.

## Categories

- Decision models: primary, weight 1

## Top 5

- Perplexity Decider v1.1 27B (Qwen3.8-27B, noncausal decision readout): 82.79186839726479
- Quyet-1.0-Large (Chinh Nguyen, Gemma-4-31B decoder, option-letter logits): 81.68880454670384
- torchcast-decision-27b (Torchcast AI, Qwen3.8-27B fine-tune): 81.56565836114741
- decisio v0.8.0 on gemma-4-31B-it (frozen, FP8 on load, one prefill per state, self-hosted): 79.56974174390604
- Kev 27B (Kev 1.0, Qwen3.8-27B): 78.93896122956639

Source: https://benchmarkheaven.com/jev-models/v1.6.1