Siri’s new reach crosses a computing boundary

Apple released Siri AI in English beta on Monday alongside its new generation of operating systems. The company describes an assistant that can use personal context, understand what is onscreen and take actions across apps. Its custom foundation models combine work performed on the device with Private Cloud Compute, using models developed in collaboration with Google’s Gemini team.[1]

Availability begins with enrollment, rather than simply installing an operating-system update. Apple’s support instructions direct eligible users to the Siri settings and the Try Siri AI option, followed by voice enrollment and a privacy review. Enrollment adds the user to a waitlist; the download begins when that wait ends. The device and Siri languages must match a supported language.[2]

That sequence matters to any assessment of the launch. An announced capability, an eligible device and an activated account describe different stages. A reader comparing experiences across devices could encounter a difference in rollout status before encountering a difference in model performance.

What runs nearby, and what does not

Apple’s current Apple Intelligence requirements also distinguish devices by the local resources they need. Its support page lists different storage requirements for different hardware, up to 14GB on some newer configurations and up to 8GB on other eligible devices. These are storage requirements for the software and models. They are not measurements of working memory, response speed or the amount of personal information sent to a server.[3]

Region rules further narrow availability. Siri AI currently supports English. In the European Union it is available on eligible Macs and Vision Pro devices, but not on iOS, iPadOS or watchOS. It does not currently work for accounts whose region is mainland China. These are Apple’s published rollout conditions as of September 16.[2]

Apple presents the on-device layer as a way to handle personal context and app access, with more capable cloud models available for complex requests. The announcement also describes a dedicated Siri app whose conversation history syncs through iCloud. Audio Intelligence features including Live Rewind and Siri Recap are scheduled for later in 2026, rather than part of this week’s initial availability.[1]

An assistant living in a phone’s interface can therefore have several execution locations. Looking local does not establish that every inference stays on the device. Nor does the presence of a cloud component mean that every request needs to send the same information. The useful privacy question concerns the path of a particular task and the data required at each step.

Apple’s Private Cloud Compute architecture is intended to constrain that remote step. In its technical account, Apple describes encrypted requests to verified server software, no privileged administrative access to user data and removal of request data after processing. It publishes software information to support external inspection. These are architectural claims about how its cloud service handles data, not a claim that the computation occurs on the user’s own hardware.[4]

The distinction is fundamental because a server performing inference must process the request. Apple’s design seeks to limit who can inspect that data and how long it exists. Its protections depend on the specified hardware, software and verification mechanisms working as described.[4]

The test is a completed action

The more interesting product question comes after activation. Suppose an assistant is asked to find the latest agenda for a meeting and create a reminder containing the right document. That is several tasks: identify the meeting, select the correct version, preserve the link and write the intended reminder. A fluent explanation can accompany an error at any of those stages.

A useful evaluation would inspect the resulting reminder and its source document. It would also distinguish finding information from changing something in an app. The second step introduces a need to understand what will happen, whether permission is appropriate and how the result can be checked. Those are practical measures of assistance that a conversational demonstration may conceal.

This week’s release opens that evaluation to eligible users as enrollment completes. Apple’s description establishes the intended reach and the published computing boundaries. Reliability across real tasks remains a separate question, to be judged by correct sources, appropriate actions and inspectable results.