Project profile

Kev

Qwen-based decision models that serve TypeSafe's /v1/systemone shape so the official Python SDK can point at a local server.

Decision ModelOpen SourcePythonApache-2.0Active

Jared Palmer GitHub Demo Documentation

Last verified: 2026-09-30

Quick facts

Project	Kev
Creator	Jared Palmer
Type	Decision Model
Language	Python
License	Apache-2.0
Uses Jev for	Local server for the System One request shape
Open source	Yes
Status	Active

What it is

Kev is a family of decision models on Qwen3.5 and Qwen3.8 bases, published by Jared Palmer under Apache-2.0. The README says pretrained weights cover 0.8B, 4B, 9B, and 27B, and that a local server accepts the System One request: state plus noul, choice, and score questions.

Why it matters

It is a public example of keeping a Jev client and changing the base URL. The model, the training data, and the calibration are Kev's, not TypeSafe's.

Relationship to Jev

Hosted Jev	Not called. Weights and training are Kev's.
Wire shape	README: POST /v1/systemone, and the TypeSafe Python SDK works against that server unchanged.
Question types	noul, choice, and score in one request. Questions share the text and cannot read each other.
Sizes	README table: Kev-0.8B, Kev-4B, Kev-9B, Kev-27B.

Architecture

State and questions
Kev checkpoint
Probability distribution
Your threshold in code

Key features

Four published sizes, from a laptop 0.8B to a 27B that the README says needs an 80 GB GPU.
Each checkpoint ships with a temperature fitted on held-out data, according to the README.
A Hugging Face Space runs Kev-4B and Kev-0.8B in the browser.

Performance and claims

Project claim

The README table lists Kev-27B at 0.848 development accuracy on sources it says the model never saw, next to a Jev figure of 0.857 on those same development sets. Kev-4B and Kev-9B are listed within four points of that Jev figure.

Independent verification

Jev AI Hub read the README and the GitHub repository record on 2026-09-30. It has not re-run the suites or the hosted Jev column.

Jev AI Hub analysis

Treat the table as the author's comparison on the author's sets. A base-URL swap checks that the client still parses. It does not check that a threshold tuned on jev-1.13.0 still means the same thing.

Limitations

The README says the Jev column in its table was run only on development sets, and that the comparison is not a controlled architecture test because Jev's training data is unknown.
Kev-27B starts from a post-trained Qwen release whose own training mix the README says is unknown.
This hub has not run the frozen eval suites.