tl;dr: Kev is an open-source family of small decision models (0.8B, 4B, 9B) built on Qwen3.5 that mimics Jev's architecture, offering yes/no, multiple-choice, and rating outputs with calibrated probabilities via a LoRA adapter and pointer head. It runs locally on CUDA, ROCm, or Apple Silicon (though slow on Macs due to missing DeltaNet kernels), is API-compatible with TypeSafe's System One SDK, and supports fine-tuning from released checkpoints. Kev-9B roughly matches Jev on accuracy but trails on calibration and knowledge-heavy tasks like MMLU-Pro (0.52 vs 0.84).