Show HN: Newt – A Swift Package for Jev-Style Decisions on Apple Core AI (github.com)

Newt, a new Swift package, has been unveiled for making decisions based on graded, multiple-choice, and yes/no questions using on-device language models in Apple's Core AI environment. Named after Isaac Newton, similar to how TypeSafe's model is named after William Stanley Jevons, Newt promises high accuracy, especially for clear-cut inputs, while still maintaining a strong reliability even in ambiguous contexts. For instance, it scored a near-certain 0.9996 probability for a customer inquiry about speaking to a human agent, significantly outperforming TypeSafe. The significance of Newt lies in its capacity to run completely on-device, which safeguards user privacy by not transmitting data externally. It employs an architecture that combines the capabilities of open language models like Qwen3 with a Swift API styled similarly to TypeSafe, enhancing accessibility for developers in the AI/ML community. Newt allows for efficient on-device querying, returning detailed outputs on probabilities, confidence, and coverage, thereby catering to diverse applications such as customer service and user support. With its seamless integration into Apple's ecosystem, Newt could streamline AI decision-making processes across various applications, proving a valuable tool for developers focused on enhancing AI capabilities.