Kev: tiny open decision models that answer without generating text

Jared Palmer's Apache-2.0 family (0.6B–8B) returns calibrated yes/no, choice and score probabilities in one forward pass — Kev-8B edges hosted Jev in-domain.

Nowline SEP 21 5:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Typed questions in, calibrated odds out

    Kev takes your app state plus a batch of typed questions and returns probabilities in a single forward pass — no text, no JSON to parse. Three question types: yes/no (probability of "yes"), multiple-choice (2–255 options with per-option confidence), and ordered score (expected level plus confidence).

  • An open stand-in for TypeSafe's Jev

    Jev is TypeSafe AI's closed, hosted "System One" model for typed decisions; Kev is the version you run yourself. It keeps the same SDK shape — POST /v1/systemone — so you can point existing Jev code at a local server and stop shipping your app state to someone else's API.

  • Three sizes, all LoRA on frozen Qwen3

    Kev-0.6B, Kev-4B and Kev-8B are each just a rank-16 LoRA adapter plus a pointer head on a frozen Qwen3 base. The 0.6B is small enough to train and serve on a MacBook, answering a batch of typed questions in a fraction of a second — small enough to sit inside a request path.

  • Kev-8B beats hosted Jev in-domain

    On the locked test set, Kev-8B scores 0.870 in-distribution versus hosted Jev's 0.845, though it trails out-of-domain (0.780 vs 0.857). Kev-4B lands 0.856/0.806. Fine-tune on your own labels and the edge lands where it matters — your domain.

  • Reproduce the whole thing for ~$95

    Palmer shipped the training recipes, dataset manifests and experiment plans; the full Qwen3.5 run reportedly cost about $95 of H100 time plus three cents of Jev API calls. Clone the repo, pull weights from the release, and you can stand up calibrated routing, moderation or scoring gates this weekend — with no per-call API bill.