Kev: small, self-hostable Jev-like decision models
Kev packages open-weight Qwen-based models for fast decision tasks: yes/no, multiple choice, and ratings, with calibrated probabilities and model sizes from 0.8B to 27B. The repo includes training code and lets users fine-tune for their own labeled examples.
The authors’ benchmark table shows strong results on some held-out datasets, but these are project-reported evaluations—not independent validation. HN commenters liked the prospect of a local, low-latency classifier for routing and agent-control tasks; others questioned how well the models generalize and whether a general decision model beats a purpose-built classifier or a full LLM.