Developer Jared Palmer published Kev, an open source family of small “decision” models in three sizes (0.8 billion, 4 billion, and 9 billion parameters), built on Alibaba’s Qwen3.5 and released on GitHub with training code and pretrained weights.
Instead of writing prose, Kev scores yes/no questions, multiple choice picks, and ratings, and it returns a probability alongside each answer, information a plain chatbot reply does not carry. It runs locally on Nvidia, AMD, or Apple Silicon hardware and copies the request format used by TypeSafe’s hosted System One product, so a team can point that company’s existing software development kit at a self-hosted Kev server instead of paying for the cloud version, according to the project’s GitHub README.
Palmer’s own benchmarks, run against a suite the README calls decision-v7, put the largest model a few points behind the hosted competitor Jev on out-of-sample accuracy and confident wrong answers, a gap the README states outright rather than obscuring. Fine-tuning scripts and a Modal cloud training path are included for teams that want to close it on their own support tickets or routing rules.
According to the project’s README on GitHub (Jared Palmer), published September 22, 2026.
