The iOS 27.3 beta sends about 2% of English Siri AI requests to a local model instead of Private Cloud Compute when your device is online. It is the first time Siri AI has routed requests this way, and the requests are mostly phone calls and music.
A request handled by the local model skips Private Cloud Compute, which runs Apple’s larger server-based models for demanding requests. Every request still starts with Skimmer, the first step in Siri AI’s on-device processing, which does most of the on-device work.
pdfu found the routing in the beta and posted it on X. He later followed up with two more posts that explain how Skimmer works and how fast it is.
The local model handles 2% of requests
pdfu puts the local share at about 2% of English requests, and the routing change is in the iOS 27.3 beta that Apple seeded to developers on November 3. The same beta already uses either an on-device or a Private Cloud Compute model to fill the Siri AI memory graph.
Skimmer matches Siri AI requests with regex
Skimmer uses regular expressions to match what you say to a tool flow. When your phrasing fits one of its patterns, it handles the request with preset actions and no model has to read it. pdfu says a request such as turn on Do Not Disturb does not need a language model, since most people phrase it the same way.
One rule in the beta covers named timers. It picks up phrases such as “can you please set a 10 minute pasta timer” and pulls out the duration and a label of one to three words.
On a Mac you can browse the rules yourself. They are stored by locale in an Enigma.json file at /System/Library/PrivateFrameworks/IntelligenceFlowPlannerSupport.framework/Versions/A/Resources/enigma_patterns/<locale>/Enigma.json.
Skimmer picks a tool and fills in its arguments without using a language model, so its speed does not depend on your device. In pdfu’s testing it calls the play tool for music in under 120ms, up to 6 times faster than the on-device or Private Cloud Compute model. A simpler tool such as flip coin takes under 60ms, up to 20 times faster than a model.
Because no model runs, Skimmer also works offline. When it cannot handle a request, Siri AI compiles the relevant context and sends your query to an on-device or Private Cloud Compute model.
Pattern priority can misroute your request
The patterns differ by language, and a rule with a higher priority can take over a request meant for another feature. One user told pdfu that Siri AI in Spanish searched for music named “the living room” when they asked it to control the living room air conditioner.
pdfu replied that in Spanish, “pon” plus up to four words goes straight to music search and outranks the Home room pattern. Adding one more word sends the request to Private Cloud Compute, which should handle it correctly.
Apple has not announced the local model routing. It could change before iOS 27.3 reaches the public in December.
