Kev: Open-source decision model family built on Qwen3.5
Original: Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
Why This Matters
Self-hostable decision models lower the barrier to structured reasoning tasks without cloud dependency.
Developer Jared Palmer has released Kev, an open-source family of small decision models built on Qwen3.5, inspired by Jev's architecture. Available in 0.8B, 4B, and 9B sizes, Kev supports yes/no, multiple-choice, and rating queries in a single request, and runs on both CUDA and Apple Silicon.
Kev is a compact family of decision models released by Jared Palmer on GitHub, currently sitting at 2.1k stars. Built on top of Qwen3.5 and modeled after the architecture described in 'Jev's Architecture Unmasked,' the project lets developers train or run their own decision models without relying on a third-party service.
The models come in three sizes — 0.8B, 4B, and 9B — and handle three question types in a single request: yes/no (noul), multiple-choice (choice), and rating (score). Questions share input text but remain isolated from each other's outputs. The 4B and 9B variants fit within a 32 GB Mac using bf16 precision, making local deployment realistic for developers on Apple Silicon.
Kev's API is designed to match TypeSafe's System One, meaning existing users of that SDK can redirect calls to a local Kev server with minimal code changes. A web playground is included for testing inputs and inspecting how option ordering affects outputs. Pretrained weights and evaluation data are provided alongside training code, and Kev-4B and Kev-0.8B are also accessible directly in the browser via Hugging Face.