Kev 1.0: Jared Palmer's open decision family gets a full-weight 27B
Jared Palmer released Kev 1.0 on 1 October 2026 (post on X, 20:33 BRT): a family of four open decision models from 0.8B to 27B, all Apache-2.0, with a new Kev-27B and an updated Kev-9B. Kev uses TypeSafe’s API, so apps built on TypeSafe’s SDK switch by changing the endpoint and model name. The server runs on GPUs and on Apple Silicon through MLX, Kev-4B runs in the browser, and new agent skills cover fine-tuning and deployment.
What changed. The new Kev-27B fine-tunes every weight of Qwen3.8-27B’s text backbone (the smaller models keep LoRA adapters) on ~146k records and 337k questions, with a 64k-token document window. The released weights blend 85% of the new fine-tune with 15% of the previous 27B. Palmer says no Jev outputs were used. Kev-9B got one more pass that lifts policy and rule reasoning from 58% to 83% in his harness.
The numbers (UnverifiedClaim). Chance-corrected generalization index on 14 public datasets excluded from training: 52.3 / 41.0 / 38.0 / 23.3 (27B / 9B / 4B / 0.8B). Kev-27B pooled ECE 0.019. Median model time 23–67 ms on Modal GPUs. Palmer is upfront about the limits: the new 27B is less accurate and more overconfident on long legal contracts than the old one, base-model exposure is unknown, and checkpoint selection was not fully blind. On the official Decision Index, the pre-1.0 weights sit at 38.48 (9B) and 34.64 (4B).
How we list it. The Kev page is refreshed for 1.0: 6.1 maturity / 7.7 capability / 26 adoption, up from 4.4 / 5.5 / 8, which dated from the 0.5B prototype.