mini Jev: an open 1.05MB decision head on frozen Qwen3 0.6B outputs agent action probabilities

AlternativeSure2891 · reddit · 2026-09-21

A developer released mini Jev, an open tiny decision model for AI agents that skips text generation entirely: it takes agent state plus available actions and directly returns action probabilities. The architecture uses a frozen Qwen3 0.6B backbone, shared KV cache, parallel candidate scoring, and a 1.05MB DecisionHead with softmax, trained on 50K synthetic decisions. Available on Hugging Face (samatv256/mini-Jev); still a research prototype but fully open for experimentation.

Original post →

More from coding & agent

coding & agent channel →