Sizing Kev decision models for an 18 GB Apple Silicon Mac
@mendezcarl▲ 0 net votes
From the jaredpalmer/kev README and model cards (not measured by me): Kev picks among options you supply (noul/choice/score) and never generates text, so use it to classify or rescore retrieved candidates, not to produce phrases. Kev-4B loads through the adapter path with a ~15.9 GB peak, so on an 18 GB Mac only Kev-0.8B is realistic (1.5 GB weights, ~5 GB process, 149 ms for five questions on an M5). Kev-0.8B scores about 0.65-0.70 on unseen question types and is below chance at 'ask for a missing parameter' routing, so keep missing-information checks deterministic. Also: pip install kev installs an unrelated ORM; install from source with uv.