Back to news
researchApple2026-11-10

Apple iOS 27.3 beta reveals a new Skimmer module routing about 2 percent of Siri English requests to on-device models

According to a leaker, Apple introduced a front-end module named Skimmer for Siri in the first developer beta of iOS and macOS 27.3, routing roughly 2 percent of English requests to on-device models while online, covering high-frequency commands such as setting timers, playing music and making calls.

A source claimed in a social media post on November 10 that in the first developer preview beta of iOS and macOS 27.3, Apple introduced a front-end module named Skimmer for Siri, routing about 2 percent of English requests to on-device completion for the first time. The source said that in the macOS 27.3 preview, while connected, roughly 2 percent of Siri English requests are routed to on-device model processing, mainly covering high-frequency and relatively fixed commands such as setting timers, playing music and placing calls, which reduces latency, saves compute and protects privacy. On the background mechanism, the source said that in the new on-device processing flow Apple introduced the Skimmer front-end module, which uses regular expression matching between user requests and preset tool flows, and once a rule matches Siri can execute the operation directly without entering a more complex AI inference step. The relevant rule files can be found in macOS system directories. This path complements the 49.99 dollar Intelligence usage pack that Apple is pushing at the same time, moving high-frequency deterministic commands to the device while turning heavy inference into a purchasable increment.

AppleSiriSkimmer端侧推理iOS 27.3本地模型