Kev is a family of small decision models built on Qwen3.5 and released by developer jaredpalmer. Recent coverage describes the models in 0.8B, 4B and 9B parameter sizes, shipped with pretrained weights, training code and evaluation data, running on CUDA and Apple Silicon. The open-source tool jevals uses Kev and similar typed decision models to replace LLM judges for agent evals and guardrails. Kev handles yes/no, multiple-choice and rating questions.
OpenAI Decisions API Confidence UnreliableTests show GPT-6 Luna, behind OpenAI's new Decisions API, is poorly calibrated: when it claimed 99% confidence across 3,600 reasoning problems, it was right only 68% of the time, highlighting AI confidence-calibration problems.
Kev: Tiny decision-model family built on Qwen3.5Developer jaredpalmer released Kev, a family of small decision models built on Qwen3.5, offered in 0.8B, 4B and 9B parameter sizes with pretrained weights, training code and evaluation data. The models handle yes/no, multiple-choice and rating questions, run on CUDA and Apple Silicon (4B and 9B fit a 32GB Mac in bf16), and expose an API matching TypeSafe System One.
jevals: replacing LLM judges with typed Jev decisions for agent evalsjevals is an open-source tool that replaces LLM judges with typed Jev decision models for agent evals and guardrails. All evals for a trace are sent as one HTTP request costing a few thousandths of a cent and returning in a few hundred milliseconds, suitable for every trace or inside the agent loop. It supports Jev, Kev, Laya, and regular chat LLMs.
The A.I. Daily