# Kev

A small family of yes/no, multiple-choice and rating decision models, built on Qwen3.5 and Qwen3.8, that can be trained, fine-tuned and served from your own hardware.

Kev is a family of small decision models built on Qwen3.5 and Qwen3.8, following the architecture behind TypeSafe's Jev models, releasing four sizes from a 0.8-billion-parameter model that runs on a laptop up to a 27-billion-parameter model for a single data-center GPU.

A single request can ask a yes/no, multiple-choice and rating question together, sharing context while each answer stays independent, with calibrated probabilities from a temperature fitted on held-out data for each checkpoint. Because its API matches TypeSafe's System One, an application built against the hosted Jev API can point at a self-hosted Kev server instead, unchanged.

Pretrained weights can be used directly, fine-tuned on a user's own labelled examples through an included coding-agent skill that runs the whole loop on Modal, or deployed to a scale-to-zero HTTPS endpoint with one command; it can also be tried directly in a browser through a hosted demo space.

Website: https://github.com/jaredpalmer/kev