Jared Palmer Ports Kev to Qwen3.5 for Roughly $95 in H100

Jared Palmer released Kev on September 20, a three-model family (0.8B, 4B and 9B) built on Qwen3.5 bases and shipped under Apache-2.0. An execution log accompanying the GitHub repo puts the whole Qwen3.5 port at roughly $95 of Modal H100 usage plus $0.03 in Jev API calls, according to runtimewire.

Kev takes a document and a set of typed questions and returns calibrated probabilities across three schemas: yes/no, multiple-choice and rating. The models target, in the repo's phrasing, "small semantic judgments buried inside software workflows: routing a support ticket, choosing an escalation path or scoring a customer's frustration." The README is explicit that "Questions share the input text but can't read each other," so one answer cannot leak into the next.

Under the hood the models use rank-16 LoRA adapters over the Qwen3.5 bases with a pointer-style readout head, trained for two epochs on roughly 10,000 examples from ten public datasets, plus 896 generated policy examples and 1,680 rule-based examples. Palmer credits Archer Hume for the underlying architecture write-up and TypeSafe for the API design; the API is a drop-in for TypeSafe's System One.

On a locked test set Kev-9B posts 0.837 accuracy against the hosted Jev reference at 0.857, per runtimewire's numbers. The repo's own table shows Kev-9B at 0.852 on new sources. The port did not come free on latency: Kev-4B on an Apple M5 took 779ms for five questions, up from 174ms for the prior Qwen3 version.

The release lands five days after TypeSafe announced Jev and its System One models, a launch we tracked as it moved through 13% of Vercel paid teams inside a day. Palmer credits Devin for the training work, and the README is blunt about provenance: "No Jev outputs were used for training."