jevseek: wiring a decision model as DeepSeek's sampler to steer every token

cephaloform · x · 2026-09-23

A clever combo hack: the author plugged jev, a decision model that can't generate text, in as DeepSeek's sampler — DeepSeek proposes the top-k next tokens, and jev picks which one to say. The hybrid is dubbed jevseek, and it writes in a distinctive style.

An interesting community experiment in decoupling token selection from token generation by combining two existing models.

Original post →

More from Fun

Fun channel →