Jev-inspired local decision model

Inspired by Jev’s reported training approach, Taro L. Saito distilled judgments from DeepSeek V4 Flash into a 4B local model using 26 hours of training on a DGX Spark. The author reports about 22 milliseconds per decision and performance above the teacher’s instant mode. This is a separate local-model experiment, not a demo running Jev. Via DAIR.AI Jev Field Notes (research paraphrase, snapshot 2026-09-20).