Apple's Siri AI integrates local processing for high-frequency english commands in iOS/macOS preview

In the initial developer preview of iOS/macOS 27.3, Apple has introduced a preliminary module named Skimmer for its Siri AI. This feature routes approximately 2% of English requests to a local model for processing when the device is connected to the internet. This local handling is primarily reserved for routine, structured commands such as setting timers, playing music, and making phone calls. The goal of this implementation is to reduce latency, conserve computational resources, and enhance user privacy. The mechanism involves Skimmer acting as a front-end module that uses regular expressions to match user requests against predefined tool workflows, allowing Siri to execute actions without engaging the full AI inference process. Details regarding these rules can be found within the IntelligenceFlowPlannerSupport.framework directory on macOS, specifically in the Enigma.json file. Commenter @lolrepeatlol noted that this localized behavior might explain why Siri sometimes provides more rigid responses, such as giving a mechanical forecast like 'rain may fall within an hour' when asked when the rain will stop.