Kev

Not yet independently verified. Kev is an individual developer's project, and all accuracy and calibration figures come from the author's own evaluation harness; RuntimeWire (21 September 2026) reported them without independent testing. We will update this when it can be confirmed, and remove this note.

Kev is a family of small open-weight decision models released by Jared Palmer as a personal open-source project. Kev takes a piece of text or application state and answers several typed questions about it in one request: yes/no probabilities, multiple-choice picks with a probability for each option, and ratings on an ordered scale. It copies the System One API of TypeSafe's hosted Jev service, so TypeSafe's Python SDK can be pointed at a local Kev server.

The GitHub repository was created on 17 September 2026 with a 0.5-billion-parameter prototype. The current family, built on Qwen3.5 base models at 0.8B, 4B and 9B, was published as a GitHub release on 20 September; its Hugging Face repositories were created on 19 and 20 September, and the weights were updated on 21 September. Each model is a LoRA adapter with a small pointer head on the Qwen3.5 base. Code, adapters and evaluation data are under Apache 2.0. Kev runs on CUDA, ROCm and Apple Silicon; the 4B and 9B models fit a 32 GB Mac in bf16, though the author notes they are slow there, and the server handles one request at a time.

In the author's own evaluation, Kev-9B reaches 0.852 accuracy on a test set drawn from sources it was not trained on, and trails hosted Jev by 3.5 points on the development set; the author notes this is not a controlled comparison. Probabilities are calibrated with a single temperature, and confident wrong answers still occur. Training covered at most 1,024 tokens per state and question, while serving allows 8,192. Knowledge questions remain a large gap, option order can change answers, and users are told to test on their own data before choosing thresholds.