Kev (GitHub Repo)

Kev is a family of small decision models built on Qwen3.5, and the pitch is simple: run your own classifier-ish model locally instead of sending tickets off to somebody else’s server. The repo ships pretrained weights, training code, and evaluation data, with 0.8B, 4B, and 9B versions on offer. It also matches TypeSafe’s System One API, so existing Python code can point at a local Kev server with little fuss.

The setup is more interesting than the headline size numbers. Kev can answer yes/no, multiple-choice, and rating questions in the same request, and the questions share the same input text without reading each other’s answers. That matters because the whole point here is structured judgment: not just a label, but the probability distribution behind it. The repo even has a web playground for testing option order, plus a browser demo on Hugging Face for Kev-4B and Kev-0.8B.

The models are built on Qwen3.5 bases, and the source says Kev-4B and Kev-9B fit on a 32 GB Mac. It runs on CUDA, ROCm, and Apple Silicon via MLX, with different serving paths depending on the hardware. On Apple systems, the code uses MLX because PyTorch doesn’t have the needed DeltaNet kernels there. The repo also gives a quick-start path with Python 3.12 or 3.13, uv, and a local server on port 8009.

The evaluation section is where Kev tries to separate itself from a toy demo. Kev-9B is the strongest of the three current models, and the repo says it trails Jev by 3.5 points on the new-source development set while reaching 0.852 on the test set. All three current models were updated on 2026-09-21 with a second training pass on generated examples, which improved the test numbers a bit. Kev also stores calibrated probabilities by default, with a temperature fitted per checkpoint, so it can say how sure it is without changing the answer itself.

There’s a lot of careful engineering here, from cached state prefixes to exact question isolation on the recurrent Qwen3.5 path. That’s the real story: Kev is not trying to be a chatbot wearing a spreadsheet costume. It’s trying to be a local decision engine that knows when to shut up and give a probability instead.