iOS 27 Officially Launches Siri AI: 20B On-Device Model Loads 1B to 4B Parameters per Prompt

Apple released iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27 on September 14. With the exception of tvOS, the new Siri AI is a central theme across these releases. It can answer questions using on-screen content and personal context from sources such as Mail and Messages, and invoke actions in third-party apps through App Intents. Developers therefore need to test more than text responses: they must also validate entity resolution, parameter defaults, permission confirmations, and action rollback.

The underlying AFM 3 family includes two on-device models: the dense, 3B-parameter AFM 3 Core and AFM 3 Core Advanced, which has 20B total parameters and activates roughly 1B to 4B for each request. The latter uses Instruction-Following Pruning: the complete weights remain in NAND, while the system first determines the required subnetwork from the entire prompt and then loads the relevant weights into DRAM. This differs from conventional MoE architectures, which reroute at every token. The approach avoids the bandwidth bottleneck of swapping weights from flash storage token by token, although latency, power consumption, and model availability still depend on the device tier.

More complex reasoning and agentic tool use can be handed off to AFM 3 Cloud Pro running on Private Cloud Compute. Consequently, saying that “Siri runs on-device” does not guarantee that every request remains on the phone. The initial Siri AI rollout is still in beta and begins in English. Although iOS 27 supports iPhone 11 and later, Siri AI requires a supported device—an iPhone 15 Pro, iPhone 15 Pro Max, or iPhone 16 or later. Engineering teams should separately measure local and cloud execution paths and build regression tests around the App Intents, Search, and CarPlay edge cases that remain documented in the final release notes.