Mystery model Jev likely ~30B params with 300ms latency, relational reasoning scores 0%

paraschopra · x · 2026-09-19

paraschopra benchmarked the mysterious Jev against Qwen3-4B and a 400M-param Laya. Comparing MMLU and other results, he suspects Jev is in the 30B range, with 300ms latency matching a 30B model on a good GPU. Notably, Jev scored 0% on a relational choice benchmark, suggesting both questions and options are processed in parallel — a standard modern model with fast-inference tradeoffs.

Related event: Mystery probability-only model Jev sparks flurry of community benchmarks(7 posts)→

Original post →

More from Models

Models channel →