Kev: open decision models let AI agents choose without generating text

Developer Jared Palmer released Kev, a family of open decision models built on Qwen 3.5 with 0.8B, 4B and 9B parameters. Kev is prefill-only: it scores candidates in a single forward pass via a pointer head, skipping autoregressive decoding. Kev-9B reached 83.7% accuracy on the project's locked out-of-domain test.

• Kev is released under Apache 2.0 with weights, training code and evaluation tooling
• It supports three decision types: Noul (yes/no), Choice and Score
• The 4B model processed three questions in 277 ms in bf16 on an M5
• Kev-9B scored 83.7% accuracy on the locked out-of-domain test