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.
More from coding & agent
- Code review classification saves 122k tokens, author says even 10% savings is a free win — draginol · 2026-09-21
- Microsoft open-sources IQ Solution Accelerator unifying enterprise data, knowledge and decisions — adnan_hashmi · 2026-09-21
- Build an AI team around Claude: role-based subagents instead of one chat — Aiden_Tech_Ai · 2026-09-21
- Thorsten Ball: code review and unit tests will die in the AI era — zetalyrae · 2026-09-21
- "Fix It by 6am, No Mistakes": The Sleep-On-It Agent Prompt Gag — BLUECOW009 · 2026-09-21
- Industry consensus: agent harnesses will specialize per task, moat lives in the data layer — amankhan · 2026-09-21