Jared Palmer released Kev, a family of small decision models. They are built on Qwen3.5 and follow the Jev architecture. The models handle yes/no, multiple-choice, and rating questions in one request. They run on CUDA and Apple Silicon. The 4B and 9B versions fit on a 32 GB Mac. A web playground lets users test inputs and option order effects. Kev-0.8B, 4B, and 9B are available with pretrained weights. The API matches TypeSafe’s System One for compatibility. Kev-9B outperforms its predecessor by 7.3 points on new-source tests. All weights are published on Hugging Face and GitHub.
What changed
Previous Kev models used Qwen3 bases instead of Qwen3.5.
What it unlocks
Running small decision models locally for yes/no, multiple-choice, and rating tasks with shared input text.
- 0.8B, 4B, and 9B parameter models
- Kev-9B 7.3 points ahead of Kev-8B on test set
- Brier score 0.08 lower for Kev-9B vs Kev-8B
What you need to act on it
- Python 3.12+
- uv package manager
- CUDA or Apple Silicon hardware
- 32 GB RAM for 4B and 9B models
Sources