Kev is a family of small decision models built on Qwen3.5, based on Jev's Architecture Unmasked. You can use pretrained weights or train your own. The API matches Type Safe's System One, so you can point their Python SDK at your local server. Available in 0.8B, 4B, and 9B sizes with training code and evaluation data. Supports yes/no (noul), multiple-choice (choice), and rating (score) questions in the same request. Questions share input text but can't read each other. Runs on CUDA and Apple Silicon; 4B and 9B models fit a 32 GB Mac using bf16.

Quick Start requires Python 3.12+ and uv. Clone the repo and run the 4B model locally. The first run downloads the adapter and base model. You can also specify a local checkpoint or Hub revision.

A web playground lets you try inputs and check how option order affects answers. Includes presets for testing question isolation and a chess demo where the board is input and legal moves are Choice options.

Models start with Kev-4B. Use Kev-9B for higher accuracy and capacity.