# Kev Kev is a family of small open-source decision models by Jared Palmer (of [[Vercel]]) that you can train and run yourself. It's the most serious attempt so far at an open [[Jev]]. It started as Kev-0.5B on Qwen2.5-0.5B (small enough to train on a MacBook Pro), then grew within days to sizes up to 9B on [[Qwen]]3.5. Each checkpoint is a base model plus a rank-16 LoRA adapter and a small pointer head, following the architecture Archer Hume inferred for Jev. You send one document (the state) and a set of typed questions, and get a probability distribution per question in a single forward pass. Two details make it practical: - **Drop-in API.** It implements the TypeSafe System One API, so the official SDKs can point at a local Kev server without code changes - **Permissive license.** Apache-2.0, like the Qwen base models The numbers are close to Jev but not there yet: on held-out data, Kev-9B reaches 0.852 accuracy versus 0.857 for Jev in the repo's benchmarks, while out of domain the 8B variant scored 79.6% against 85.7%. Bigger variants are also better calibrated. For me, Kev is the one to watch if you need decisions on your own hardware (privacy, cost at scale, or no waitlist). ## References - [Kev on GitHub](https://github.com/jaredpalmer/kev) - [kev-0.5b on Hugging Face](https://huggingface.co/jaredpalmer/kev-0.5b) - [Jared Palmer's announcement of the larger Kev family](https://x.com/jaredpalmer/status/2101715352258232539) ## Related - [[Jev]] - [[System One Models]] - [[SemIf]] - [[Jevlike]] - [[AI Open Weight Models]]