Kev is a family of small decision models built on Qwen3.5 that you can train and run locally.
The tool addresses the need for lightweight, self-hosted decision models that can handle multiple question types in a single request. It builds on the Jev architecture and provides models in three sizes. The approach allows questions to share input context while remaining isolated from each other, and supports yes/no, multiple-choice, and rating questions simultaneously. The tool runs across different hardware platforms including CUDA, ROCm, and Apple Silicon, with the larger models fitting within 32 GB of memory on Mac systems using bf16 precision.
Kev suits teams wanting to run decision models without external API dependencies or the overhead of larger language models. The pretrained weights are available for immediate use, or you can train your own using the provided training code and evaluation data. The API matches TypeSafe's System One, allowing you to point their Python SDK at a local Kev server. A web playground lets you test inputs interactively and observe how option ordering affects model responses, available both locally and in a browser-based demo without installation.
The project maintains frozen evaluation suites for consistent benchmarking and publishes a research log documenting development decisions. Training code and evaluation data are included alongside the model weights, enabling reproducible experimentation. The tool provides serving performance documentation specific to different hardware configurations, helping users understand what to expect on their target platform.
