GitHub radar
kev: Open-Source Decision Model That Runs on Your Mac
kev is a small open-source decision model — a local alternative to TypeSafe Jev — that reads a document once and answers classification questions in a single forward pass, returning probabilities instead of generated text.
kev is an open-source replica of TypeSafe Jev, a commercial decision model that classifies or routes text by answering typed questions — which department handles this ticket, how frustrated is the customer, should this be escalated. Internally it is a LoRA adapter built on Qwen base models from 0.5B to 8B parameters. The model encodes the document once and then answers all questions in one prefill pass under a block-causal mask, returning probabilities rather than generated text. It supports three question types: yes/no (noul), multiple choice, and ordered score. The API is fully compatible with the official TypeSafe SDK, so any code that calls Jev today works against a local kev server with a base_url change. On out-of-domain data, kev-8b scores 0.774 compared to Jev's 0.857.
Why a vibe-coder should care
Useful for anyone building automations with routing or classification — support ticket handling, content moderation, yes/no decisions in a pipeline — without paying an API on every call. Quality is honest: on simple, repetitive tasks it is enough; for complex or unfamiliar inputs, the commercial Jev is more accurate.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up kev-4b from https://github.com/jaredpalmer/kev — install dependencies per the README, download the weights from Hugging Face, start the local server, and show me a working example that routes customer support tickets to the right department. Ask me if you need a Hugging Face token to download the weights.
kev-4b needs a Mac with an M-series chip and 32 GB of RAM. The 0.5B version works with less memory, but accuracy is lower.
Open on GitHub▌ More finds