Kev: Small Decision Models You Can Train Yourself

🔥 What happened Jared Palmer dropped Kev — a family of small decision models (0.8B, 4B, 9B) built on Qwen3.5 that you can train and run locally. The API mirrors TypeSafe's System One, runs on CUDA, ROCm, and Apple Silicon, and Kev-4B fits on a 32 GB Mac. 💡 Why it matters Instead of a single label, the models return probabilities for yes/no, multiple-choice, and score questions in one request. Kev-9B hits 0.852 on new sources — just 3.5 points behind Jev, whose training data is unknown. A second training pass on 2026-09-21 lifted Kev-9B from 0.837 to 0.852. ⚡ Our take Finally, decision models that don't hide behind an API wall. If you need calibrated classification, grab Kev-4B today — the 0.8B variant is more of a toy.