Jared Palmer Releases Open-Source Kev Decision Models

Jared Palmer has expanded Kev, an open-source implementation of TypeSafe's Jev decision-model pattern, into three new sizes: 0.6B, 4B, and 8B. Built on a Qwen3 backbone, these models accept a document alongside a set of typed questions and return probability distributions in a single forward pass. By treating workflows like ticket routing, document classification, and risk scoring as classification tasks, Kev completely bypasses the latency and parsing overhead typically associated with prompting generative models for JSON outputs.

Under the hood, Kev attaches a LoRA adapter and a pointer head to the Qwen3 base. It packs the document and questions into a single token sequence, using a block-causal attention mask so questions can attend to the shared state in parallel without influencing each other. The pointer head then compares each option's hidden state to a special decide token, using softmax to output probability distributions. The architecture supports choice schemas ranging from 2 to 255 options, as well as score schemas like a 1-to-5 rating.

In terms of performance, Kev-8B achieved a 79.6% out-of-domain score on held-out data, compared to Jev's 85.7%. The Kev-4B model runs in bf16 on a 32GB Mac, processing five questions in approximately 300 milliseconds, or just 40 milliseconds on an Nvidia H100 GPU. Training these models is highly accessible, requiring only 40 minutes for the 4B version and 83 minutes for the 8B version on a single H100.

For developers, Kev acts as a drop-in replacement for TypeSafe's System One API and integrates seamlessly with their existing SDK. When multiple requests reuse the same document state, a key-value cache can speed up computation by 2 to 2.5 times. Released under the Apache 2.0 license, the full code and weights are now publicly available on GitHub.