Kev 4B

Kev 4B is an open-weights decision model from the Kev family, a LoRA fine-tune of Alibaba's Qwen3.5-4B-Base with a pointer head that reads the answer from a single forward pass. You send a state (text, a JSON object or an array of text) and named questions, and it returns typed answers: the probability that a statement is true, a choice among options you define, or a position on your own rubric, each with a full probability distribution. It generates no text, so answers cannot fall outside the options given. Kev speaks the same request and response format as TypeSafe's System One API, so code written for Jev can switch models by changing the model name. Each checkpoint applies a fitted temperature so its probabilities are calibrated by default. Its authors publish evaluations against Jev: Kev trails Jev on most public suites and matches or beats it on some routing tasks, and accuracy drops on documents longer than its 384-token training context. Use it for atomic decisions composed in code: ticket routing, intent detection, guardrail checks and rubric scoring in place of an LLM judge, with the probability deciding what runs automatically and what escalates. Kev 4B is published by Jared Palmer at https://huggingface.co/jaredpalmer/kev-4b and is built on Alibaba's Qwen3.5-4B-Base (https://huggingface.co/Qwen/Qwen3.5-4B-Base).