At WWDC 2026 Apple unveiled a completely rebuilt Siri, powered by a new on-device AI model that breaks the traditional memory bottleneck. The AFM 3 Core Advanced architecture stores its 20-billion-parameter weight set in NAND flash instead of DRAM, using a technique called Instruction-Following Pruning (IFP). The system routes experts once per prompt, activating between 1 and 4 billion parameters depending on task complexity, while always keeping a shared set of experts in working memory.
Why this shift matters
On-device AI agents have long been constrained by DRAM capacity, forcing architects to choose between capable cloud-dependent models and limited local ones. Apple's approach removes that trade-off: simpler requests stay on the device, ensuring privacy and low latency, while complex agentic tasks route to the AFM 3 Cloud Pro server model running on Nvidia GPUs in Google Cloud, still within Apple's Private Cloud Compute boundary. This matters especially in regulated industries where data residency and inference documentation are critical. As recent cybersecurity incidents like the critical Linux bug and the Tchap breach show, on-device processing shrinks the attack surface significantly.
macOS Golden Gate and iOS 27: concrete updates
Beyond Siri AI, macOS Golden Gate delivers major Liquid Glass design refinements: a transparency slider, uniform toolbars, edge-to-edge sidebars, and colored icons. iOS 27 introduces 12 new features including AI-enhanced Flyover in Maps, custom duration location sharing, bill splitting via Apple Cash and Apple Intelligence, and custom pass creation from physical cards using the camera. ICloud+ subscribers get higher daily usage limits for server-dependent AI features like image generation, though the entry-level tier (0.99 USD monthly) is excluded. New AirPods beta firmware supports a revamped interface and custom EQ, tightly integrated with the new Siri. The technical report with full benchmarks is due later this summer, addressing gaps in power, thermal, and compliance documentation noted by developers.
Sponsored Protocol
