Kev: Self-Hostable Decision Models on Qwen3.5 That Return Calibrated Odds

Kev is an openly published family of small “decision models” — 0.8B, 4B, and 9B parameters — fine-tuned on Qwen3.5 bases and modeled on the closed ‘Jev’ architecture. Rather than generating free text, each model answers structured questions about an input: yes/no, multiple-choice, and 0-to-N rating, with several questions batched into one request while kept isolated from one another. Crucially, it returns full probability distributions instead of a single label, so a support ticket touching returns, shipping, and billing surfaces as split odds across all three rather than one forced pick. The project ships training code, evaluation data, pretrained weights with SHA-256 checksums, and a web playground, and it deliberately mirrors the ‘TypeSafe System One’ API so an existing Python SDK can be pointed at a local server. It runs on CUDA or Apple Silicon, with the 4B and 9B fitting a 32GB Mac in bf16.

The release leans hard on measurement honesty. Benchmarks report Brier scores across held-out training sources and genuinely unseen ‘new sources,’ and the authors concede their comparison against Jev isn’t controlled because Jev’s training data is unknown — Kev-9B trails it by 3.5 points on the new-source dev set. A same-day second training pass on synthetic policy and evidence-removed cases produced small but statistically bounded gains, and the older Qwen3-based weights remain available specifically because swapping only the base makes generation-to-generation gains a clean, controlled comparison.

Two optional knobs adjust probabilities without changing answers: a fitted temperature that roughly halves calibration error and confident-error rates on new data, and a flag that injects day-count arithmetic between dates the model can’t reliably compute itself, sharply improving deadline-policy questions. The overall pitch is a local, auditable alternative to a hosted classification API, aimed at teams that want calibrated confidence they can inspect rather than opaque labels from a black box.