Independent Local Model Jev-Style 2B v4 Brings Jev-Format Decisions to Consumer Hardware
Maintainer Chaoliang Yan released Jev-Style 2B v4 for local execution, trained partly on public Jev API outputs and highlighting both edge demand and distillation questions.
Developer demand to run Jev-style typed classification completely offline reached a new milestone in late October with the release of Jev-Style 2B v4, an independent local decision model maintained by Chaoliang Yan with no affiliation to TypeSafe AI.
Edge deployment without API round trips
Designed to run on ordinary developer laptops and modest single-GPU servers, the 2-billion-parameter model accepts the same bounded question format popularized by TypeSafe’s System One API. For local agent loops, desktop apps, and privacy-sensitive document workflows, eliminating network round trips lets software run dozens of gating and routing checks locally at zero marginal API cost.
The distillation question across decision APIs
Notably, Yan disclosed that Jev-Style 2B v4 was trained in part using calibrated answers collected from Jev’s public API. That detail underscores a structural dynamic unique to decision models: because hosted System One endpoints return explicit probability distributions over options at pennies per million tokens, distilling their decision boundaries into compact local weights is dramatically cheaper than distilling long chain-of-thought reasoning traces. How hosted providers balance open developer access against distillation terms is quickly becoming a key policy question for the category.