After years of postponements that reportedly cost some employees their jobs, Apple has finally shipped a completely rebuilt Siri. The assistant, now marketed as "Siri AI," is available today as a beta and represents the deepest overhaul of the product since it first launched. The most surprising detail is not the interface or the new voices — it is what sits underneath: Google's Gemini models.
The release is a notable moment for Apple, which has spent much of the past several years positioning itself as a company that keeps as much computation as possible on the user's own hardware. Handing the foundation of its flagship assistant to a rival's model family is a pragmatic admission that building frontier-scale language models in-house takes longer than the market was willing to wait.
The Gemini foundation behind the new Siri
According to Apple, the new Siri runs partly on the device and partly through Private Cloud Compute, the company's privacy-oriented server infrastructure. Apple says users' personal data is neither stored nor made accessible to the company itself — a claim that has become central to how Apple frames its AI story. The hybrid design means simpler requests can be handled locally, while heavier reasoning tasks are offloaded to the cloud.
Whether that split holds up under scrutiny is an open question, but it reflects a broader industry pattern: capable assistants increasingly require a mix of local silicon and remote compute. Apple's answer is to keep the cloud tier proprietary and contractually constrained rather than routing everything to a third-party API.
- On-device processing for lightweight tasks
- Private Cloud Compute for more demanding requests
- Personal data neither stored nor exposed to Apple, per the company
- Gemini models supply the underlying language capability
What the rebuilt assistant can actually do
The headline capability is personal context. Apple says Siri can now draw on information from messages, emails and photos, read what is currently on the screen, and pull in results from the web. That combination is meant to move the assistant past single-command interactions and into multi-step tasks.
System-wide app actions extend the reach further. Siri can draft an email or edit and share a photograph on the user's behalf. Apple's own example involves asking for a family recipe that was mentioned in a Messages conversation, then having Siri move the ingredients into the Reminders app — a chain of actions that would previously have required several manual steps.
Third-party integration is part of the pitch as well. Siri works with apps including WhatsApp and Audible at launch, with Outlook, Notability and Tripsy planned to follow. There are also several ways to summon the assistant.
- The "Hey Siri" wake phrase
- The side button
- The Dynamic Island
- Spotlight on iPad and Mac
A dedicated Siri app syncs chat histories across devices via iCloud, and Apple is adding new voices alongside a reworked dictation feature that the company says runs on its on-device AFM Core Advanced model.
A gradual rollout, with the EU and China left out
For now, Siri AI is English-only and labeled a beta. Apple says support for French, Japanese, Korean, Portuguese and Spanish will arrive next month. The features are tied to the 2027 software cycle, appearing across iOS 27, iPadOS 27, macOS 27, watchOS 27 and visionOS 27.
The more consequential limitation is geographic. The service will not launch in the European Union or China at this stage. Regulatory friction in both markets has repeatedly shaped how large platform owners stagger AI feature releases, and Apple's decision keeps the most ambitious parts of its assistant out of two major regions for the time being. Users there will have to wait for a timeline Apple has not yet specified.
Beyond Siri: Apple Intelligence spreads across the system
Siri is only part of the story. Apple is distributing AI features throughout its operating systems, touching apps that go well beyond the assistant itself. The Photos app gains tools including Spatial Reframing, Extend and a reworked Clean Up function.
Image Playground can now generate photorealistic images, a capability Apple says runs through Private Cloud Compute. The company has also indicated that the feature will support the SynthID standard, a watermarking approach designed to signal that an image was synthetically produced — a detail that matters as pressure grows on platforms to label AI-generated media.
- Photos: Spatial Reframing, Extend and a rebuilt Clean Up
- Image Playground: photorealistic generation via Private Cloud Compute
- Watermarking: planned SynthID support
Early impressions: practical wins, lingering confusion
Early testers have generally praised the practical improvements. The assistant appears more useful in everyday situations, handling context-aware requests and cross-app actions that the old Siri could not. That is a meaningful shift for a product that had become shorthand for unfulfilled promises in the voice assistant category.
The praise is not unconditional. Reviewers report that misunderstandings still occur, suggesting the new architecture has not eliminated the failures that have long frustrated users. Multi-step reliability, in particular, remains the hardest problem to solve — an assistant that completes four steps correctly and fails on the fifth can be more aggravating than one that never tried.
What to watch next
The beta label gives Apple room to iterate, and the language expansion expected next month will be an early test of whether the Gemini-backed system generalizes beyond English. The bigger questions are structural: how much of the experience Apple can eventually bring back in-house, when regulators in the EU and China will see the feature, and whether the privacy guarantees around Private Cloud Compute hold up under independent examination.
For now, Apple has done something it had avoided for years — shipped the assistant it kept promising, with a competitor's technology doing much of the heavy lifting.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
