nibble
Answers an agent's multiple-choice questions fast and says how sure it is.
The target is a small model that picks among typed options in an agent pipeline, returns calibrated odds, and beats a big model on speed. It builds on Kev, an open model. A first comparison on 50 held-out prompts: Kev-0.8B on CPU scored 0.88 against 1.0 for Sonnet and ran about 1.3x faster. That sample is too small to call, and the gap is real.
The first experiment has Kev-4B compose shadcn landing pages, then a fine-tune on rubric labels on my shared 3060. It is exploratory and lives on a branch.
My earlier classifiers taught me the most: one scored 44.5% on unseen wording at 97% confidence. Calibration on new phrasing is the real test.