kev and SemIf: open-source Jev-style decision models
Two local-first projects for semantic if decisions: kev (Qwen2.5-0.5B LoRA on a MacBook) and SemIf (open-model scoring on a 3090 or in the browser).
After Jev went viral, open-source alternatives landed that you can train and run locally.
kev — Jared Palmer's tiny Jev-inspired model on Qwen2.5-0.5B. A LoRA adapter plus readout head reads a document once and answers many typed questions in one forward pass. Supports noul, choice, and score question types with calibrated probabilities. Implements TypeSafe's System One API, so the official SDK works with a local server. Trains in ~1h45m on Apple Silicon and serves six-question requests in ~160 ms.
SemIf — Independent semantic-if baseline from open models. Reads typed option probabilities directly from a 4B model with no JSON decoding loop. Runs on CUDA (3090-class GPU), Apple Silicon via MLX, or in a WebGPU browser demo. Formerly OpenJev; not affiliated with Jev or TypeSafe.
Both are good starting points if you want Jev-style runtime decisions without a closed waitlist.