Self-host jevos

jevos is a small decision model plus a single server binary that gives probabilistic answers to yes/no questions, multiple-choice routing and scored ratings, and it speaks the TypeSafe Jev wire format so existing Jev SDK code keeps running locally. It targets rule-driven decisions (refund eligibility, ticket routing, tone scoring) done entirely on your own hardware: inference is CPU-only, in roughly 28-130 ms per request, with an 8,192-token context and about 1 GB of memory. The server exposes POST /v1/systemone and a /health endpoint, can answer one-shot decisions from a request file without running as a service, and accepts an optional bearer key; the model also ships as GGUF for llama.cpp and as an OpenVINO INT8 build. Installation means downloading the release archive for Windows, Linux or macOS (Apple silicon), placing the model beside the binary, and launching ./jev serve. The project is MIT-licensed; in the author's own benchmarks it is much faster and free versus Jev's cloud API but less accurate on the harder tasks, so they advise validating it against your own questions first, and the score output type is still marked early.