AI-generated editorial illustration; not a documentary photograph.
Developer Jared Palmer has released a new generation of Kev, a family of specialized decision models designed to make bounded AI agent choices without generating text. Built on Qwen 3.5, the release includes 0.8 billion, 4 billion and 9 billion parameter variants under Apache 2.0, alongside training code and evaluation tools.
One pass instead of decoding
Kev processes context, questions and candidate answers in a single forward pass, then obtains decisions from a pointer head rather than running an autoregressive decoding loop. It supports yes-or-no decisions through Noul, candidate selection through Choice and ordered levels through Score. Returned probabilities are limited to the supplied candidates, preventing the model from introducing an unlisted option, although it can still select the wrong one.
A narrower role in agent stacks
The models are intended for repetitive, constrained steps such as tool routing, ranking, safety checks and escalation. Multiple questions can be evaluated against the same state in one pass, with a block-causal attention mask separating them. Palmer's documentation reports that Kev-4B processed three questions in 277 milliseconds using bf16 on an M5, but no controlled comparison with equivalent Qwen generation was provided, so the practical speed advantage remains unquantified.
Accuracy and calibration limits
According to Palmer's model card, Kev-9B achieved 83.7% accuracy on a locked out-of-domain test. This is a developer-reported result, and the evaluations identify calibration drift on unseen source distributions. That matters when an agent uses probability thresholds to execute actions or escalate decisions, because a high score does not guarantee a correct choice.
Fine-tuning also reduced some inherited general-knowledge and arithmetic performance, particularly in smaller models. Kev is therefore positioned as a specialized component beside a general-purpose reasoning model, not as its replacement. Applications making only occasional bounded choices may find constrained decoding with an existing model simpler than operating another model.
Compatibility with System One
Kev implements the request and response format of TypeSafe's System One API, allowing applications built for that interface to target a local Kev server. TypeSafe previously introduced Jev using the same Noul, Choice and Score primitives, but Jev's weights and training data are not public, limiting direct comparisons between the two implementations.
Source: Your AI agent is burning tokens on choices that don’t need words
SIGNAL TITANS
Leave a Reply