Kev: tiny Jev-like decision models (0.8B/4B/9B) on Qwen3.5 you can train and run locally - the 9B fits a 32GB Mac

Kev is a new family of small, Jev-architecture decision models (0.8B, 4B, 9B) built on Qwen3.5. These Apache-2.0 licensed models can be trained and run locally, with the 9B version fitting on a 32GB Mac in bf16. Kev is API-compatible with TypeSafe's System One, allowing its Python SDK to point to a local server without changes. A web playground demonstrates how option order can bias answers, a common issue for LLM-as-judge applications.

Unlike other judge models, Kev offers Apache-2.0 licensed models with training code and frozen eval suites, and its 9B version fits on a 32GB Mac.