Small decision models
ossA month of research on open Jev-style decision models: a decision router for my agent harness, three small public models, and Kev on llama.cpp.
TypeSafe AI’s Jev is a closed “System One” model: you send it a state and a list of typed questions, and it returns a probability for every option instead of generating text. Between 23 September and 12 October 2026 I worked on its gap with open models on a single 16 GB RTX 4070 Ti Super, measured by coverage at 5% error: the largest share of questions a model can answer while the error rate among the answered ones stays at or under 5%.
On the same 1,046-question general set, every 4B to 9B model I scored lands between 0.550 and 0.557: Kev-4B, decider-4b, Cloudflare’s Clef-flash, and my llama.cpp port of Kev-4B. Every attempt to get a smaller model there (a raw MoE screen, Kev’s training recipe on Granite-H-350M, distilling Kev-4B into Kev-0.8B) finished at or below Kev-0.8B’s 0.276. Along the way I built verdict, a decision router that serves these models to my agent harness, an adversarial panel for shell-command risk, and three small public models: granite-4.0-h-350m-decision, qwen3.5-0.8b-decision-distill-4b and qwen3.5-0.8b-decision-gold-control.
License: the models are research and non-commercial, because their training data includes ANLI, RACE, ag_news and Yelp.