The Verge attended a live, on‑the‑record technical session on Monday in San Jose, following Apple’s WWDC 2026 keynote. Apple’s senior vice president of software engineering, Craig Federighi, led the discussion with engineers from his team, detailing the architecture that powers the company’s latest Apple Intelligence features.
The deep‑dive came hours after Apple introduced a suite of AI‑driven tools across iOS, macOS, and watchOS. Federighi explained that the new framework is built on a unified neural‑processing stack, allowing developers to tap on‑device models without sacrificing performance. The approach emphasizes privacy, keeping user data local while still delivering sophisticated language and vision capabilities.
Federighi described the stack as a three‑layer system: a low‑level hardware accelerator, a middle‑level runtime, and a high‑level API for app developers. „Our accelerator runs at 2 teraflops per watt, which means we can execute large models without draining the battery,” he said. The runtime automatically selects the optimal execution path, balancing speed and energy use. The API, called Apple Intelligence Kit, offers pre‑trained models for tasks such as speech transcription, image tagging, and contextual suggestions.
Engineers highlighted how the architecture integrates with existing frameworks like Core ML. By reusing familiar tools, developers can migrate existing models to the new stack with minimal code changes. Federighi noted that the system also supports third‑party model formats, expanding the ecosystem beyond Apple‑only solutions. The live blog captured dozens of technical questions, with Federighi providing concrete code snippets and performance metrics.
Analysts are watching to see whether Apple’s emphasis on on‑device intelligence will pressure rivals to prioritize privacy‑first AI. The company’s new stack could reduce reliance on cloud‑based inference, a move that may appeal to enterprise customers wary of data exposure. Federighi argued that „privacy is not a trade‑off; it’s a feature,” suggesting that Apple aims to set a new industry standard.
Critics point out that Apple’s hardware advantage may limit the reach of its AI platform to newer devices, potentially leaving older hardware behind. However, Federighi assured that the runtime can gracefully degrade performance on legacy chips, ensuring a baseline experience across the product line. The session concluded with a preview of upcoming developer tools that will streamline model training directly on Mac machines.
Apple’s new architecture signals a deeper integration of AI into everyday workflows. If the on‑device model delivers on its promises, developers could build richer, more responsive apps while preserving user privacy. The tech community will likely gauge success by the speed at which third‑party developers adopt Apple Intelligence Kit and the real‑world performance of the announced features.
What is the main advantage of Apple’s on‑device AI architecture? It processes data locally, reducing latency and protecting user privacy while maintaining high performance through a dedicated hardware accelerator.
Can developers use existing Core ML models with the new system? Yes, the architecture is designed to be backward compatible, allowing developers to migrate Core ML models with minimal changes.
Will older Apple devices benefit from the new AI capabilities? Federighi said the runtime adapts to older hardware, offering reduced but functional AI features on legacy devices.