kev: Small decision-making models based on the Jev architecture running on 32GB Macs

This is a series of small decision-making models that reproduce the Jev architecture based on Qwen3.5. It is available in three sizes: 0.8B, 4B, and 9B, and can be used directly with pre-trained weights or trained on your own data. It is compatible with TypeSafe's System One API, allowing you to connect existing Python SDKs and run it on a local server.

It simultaneously handles yes/no (noul), multiple-choice (choice), and score (score) questions for a single input text. Each question is evaluated independently to avoid mutual interference, and returns probability values rather than single labels to gauge confidence. Probabilities are calibrated by default, managing both prediction accuracy and calibration error for new data sources.

It supports both CUDA and Apple Silicon. The 4B and 9B models can run on 32GB memory Macs with bf16 precision. Through the web playground, you can visually verify changes in answers based on option order or the effect of question isolation, and directly test the model's behavior through various examples such as a chess demo.