Apple’s Core AI framework is finally giving developers native access to on-device large language models, but the raw inference API is often too low-level for practical application building. Enter 'Newt,' a new Swift package from developer willswire that introduces 'Jev-style' decision-making patterns directly into the Core AI workflow. The project, posted to Hacker News on September 26, 2026, aims to simplify the process of routing user intents to specific AI capabilities without requiring developers to write extensive custom parsing logic.

The Problem with Raw Inference

As on-device AI becomes standard in iOS and macOS applications, developers are struggling with the gap between generating text and making actionable decisions based on that text. Core AI provides the engine, but it doesn’t offer a built-in mechanism for structured output or decision routing. Newt addresses this by providing a Swift-native abstraction that leverages structured generation techniques—often referred to in the industry as Jev-style or similar schema-constrained methods—to ensure the AI's output can be reliably mapped to application code.

Early Stage, High Potential

The package is currently in its infancy, having just been showcased on Hacker News with minimal community engagement so far (2 points, 0 comments). However, the timing is critical. With the Apple Silicon ecosystem fully embracing on-device LLMs, tools that reduce the friction of integrating these models into production Swift apps are in high demand. Newt is an early attempt to standardize how Swift developers interact with Core AI beyond simple chat completions.

Key Takeaways

  • Newt is an open-source Swift package hosted on GitHub by user willswire.
  • The tool focuses on 'Jev-style' decisions, implying structured output or deterministic routing for AI agents.
  • It targets Apple’s Core AI framework, which powers on-device LLMs in recent iOS versions.
  • The project is very new, with little to no community feedback or documentation available yet.

The Bottom Line

We need more than just chat interfaces; we need deterministic control. If Newt can actually deliver robust, type-safe decision routing for Core AI, it could become an essential dependency for any serious on-device AI app.