reflex-27b (Qwen3.8-27B)
Data: JevBench v1.6.1, published 2026-10-06
Capability 74.4 · Not ranked in the current release
Last measured: 2026-09-28 · measurement release: v1.5.0
MIT code; Apache-2.0 Qwen weights
intelligence
62.8
calibration
85.9
speed
69.2
cost
25.8
Composite (secondary)
13.2 · Not ranked on this board
Cost and measurement conditions
$0.297 per 1,000 decisions (estimated)
price floor: base-model reference price applied
p50 latency: 2.7 s
x2 + 0.15 s (assumption, not measured)
evaluator-owned Lium GPU pod (H100), offline read-only container