Apple’s iOS 27 developer beta ships without the new AI dictation feature enabled by default. The system, built on the AFM 3 Core Advanced model, promises a “major boost in accuracy” with on‑the‑fly capitalization and punctuation. It runs on‑device, meaning transcription quality doesn’t depend on network connectivity. Apple expects early adopters to manually opt in during the beta cycle this summer.

Only a select group of hardware can host the 20‑billion‑parameter model: iPhone 17 Pro, iPhone 17 Pro Max, iPhone Air, Vision Pro with M5, iPads with M4 or later sporting at least 12 GB RAM, and Macs with M3 or later meeting the same memory floor. In side‑by‑side human tests, AFM 3 Core Advanced beat the prior dictation system by a 44.7%‑to‑17.6% margin across seven quality dimensions overall performance.

Apple also bundles the model into its new customizable expressive Siri voices, another opt‑in preview in beta 1. By keeping the heavyweight engine on flash storage and employing Instruction‑Following Pruning, the company demonstrates that on‑device AI can scale without cloud reliance. Consumers with supported devices gain faster, more private transcription, while older hardware remains unchanged in the current iOS rollout today.