Kev

Small open-weight decision models released as a personal open-source project by one developer make up the Kev family. Where a chat model writes a paragraph, Kev scores questions about a piece of text or the state of an application, and it can take several of them in a single request.

The kinds of answer it gives

Answers come as yes/no probabilities, picks among several choices with a probability for each, and ratings on an ordered scale. The project mirrors the System One API of the hosted Jev service from TypeSafe, which means TypeSafe's Python SDK can be aimed at a local Kev server. That makes it a way to run the same style of typed decision locally.

Sizes and how they were built

The GitHub repository dates from 17 September 2026 and started with a prototype of 0.5 billion parameters. The present family sits on Qwen3.5 base models and comes in 0.8B, 4B and 9B sizes; it went out as a GitHub release on 20 September. The Hugging Face repositories were set up on 19 and 20 September, with a weight update on 21 September. Every model is a LoRA adapter plus a small pointer head attached to its Qwen3.5 base, and the code, adapters and evaluation data carry the Apache 2.0 license.

Hardware

It runs on CUDA, ROCm and Apple Silicon. A 32 GB Mac can hold the 4B and 9B models in bf16, although the author says they run slowly there, and the server takes only one request at a time.

Results and limits, in the author's words

Kev-9B scores 0.852 accuracy in the author's own evaluation, on a test set built from sources outside its training data, and it sits 3.5 points behind the hosted Jev on the development set; the author adds that the comparison is not controlled. A single temperature calibrates the probabilities, and confident wrong answers still happen. Training used up to 1,024 tokens for each state and question, whereas serving accepts 8,192. Knowledge questions are still a large gap, the order of options can change an answer, and users are advised to test on their own data before they set thresholds.