Apple used WWDC26 to make one thing much clearer: the new Siri AI is not simply Google Gemini wearing an Apple interface. Apple’s next assistant is built around the third generation of Apple Foundation Models, or AFM 3, with on-device intelligence still positioned as the center of the experience.
That distinction matters because Apple’s AI strategy has been easy to misunderstand. Google’s Gemini is part of Apple’s developer story, and Google has confirmed that Apple developers can securely call cloud-hosted Gemini models through Apple’s Foundation Models framework and access Gemini in Xcode. But that does not mean Siri AI is just Gemini by another name.
Apple’s own machine-learning team introduced AFM 3 as a new model family for Apple Intelligence, including on-device and server models designed for Apple products. The most advanced on-device version, AFM 3 Core Advanced, is described by Apple as a 20-billion-parameter sparse model that activates only a smaller portion of its parameters depending on the request. Apple says it is natively multimodal and optimized for its most capable Apple silicon systems.
That is the story behind Siri AI. Apple is using outside AI partnerships where they make sense, but the assistant’s foundation remains tied to Apple’s own models, silicon, privacy architecture, and operating-system integration.
AFM 3 Defines Apple’s Siri AI Direction
AFM 3 is important because it shows Apple trying to move Siri AI beyond voice commands without abandoning the privacy-first design that separates Apple from many AI competitors.
The old Siri was mostly a command system. It could set timers, call contacts, control smart-home devices, open apps, send messages, and answer simple questions. Siri AI is meant to become more contextual, more conversational, and more capable across apps. That requires a very different model layer.
Apple’s third-generation Foundation Models are designed for that shift. They are not only general chat models. They are tuned for Apple experiences: voice, dictation, personal context, app actions, system understanding, and multimodal input. That makes them more relevant to Siri than a generic chatbot model would be.
The difference is purpose. A chatbot answers a prompt. Siri AI needs to act inside the operating system. It needs to understand what the user is asking, which app is involved, what content is relevant, and what action should happen next.
That is why AFM 3 matters. It gives Apple a native model family built for the device environment where Siri actually lives.
On-Device Intelligence Remains Apple’s First Answer
Apple’s AI strategy starts on the device whenever possible. That remains one of the biggest differences between Apple Intelligence and many cloud-first AI products.
On-device AI is faster for many tasks, works with tighter privacy boundaries, and lets Apple use its hardware advantage. iPhone, iPad, and Mac are not just screens for remote intelligence. They are AI-capable machines with Apple silicon, Neural Engine hardware, memory architecture, and system-level software integration.
AFM 3 Core Advanced makes that strategy more visible. Apple says the model is its most powerful on-device model and is optimized for its most capable Apple silicon systems. Because it uses a sparse architecture, the full 20-billion-parameter model does not need to activate all parameters for every request. Apple says it activates between 1 billion and 4 billion parameters depending on the task.
That design is useful because mobile and personal devices have strict limits. An iPhone cannot behave like a massive cloud data center. It has battery, heat, memory, and performance constraints. Apple’s model work is therefore not only about raw size. It is about fitting intelligence into hardware people already carry.
For Siri AI, that is essential. A personal assistant cannot feel native if every interaction depends on a remote server.
AFM Core Advanced Creates a Hardware Divide
AFM 3 also shows that Apple Intelligence is becoming more tied to hardware capability. Apple says AFM 3 Core Advanced is unlocked by and optimized for its most capable Apple silicon systems. Early reporting around WWDC26 has already pointed to the most advanced Siri AI features being limited to the newest iPhone hardware with more memory.
That could become controversial. Apple has always tied some new features to newer chips, but AI makes the gap more visible. A user may have a recent iPhone that supports Apple Intelligence, yet not receive every advanced Siri AI capability. Voice expressiveness, higher-accuracy dictation, and other heavier on-device tasks may require more RAM and stronger local processing.
This is not unusual in AI. Model performance depends on memory, compute, bandwidth, and thermal design. But Apple will need to explain the divide carefully. Users are more likely to accept feature limits when they understand that the restriction is technical rather than arbitrary.
The strategic side is clear. Apple’s most advanced on-device AI becomes another reason to buy newer iPhones and Macs. The privacy side is also clear. If Apple wants more intelligence to happen locally, it needs hardware powerful enough to run those models well.
Gemini Is a Developer Option, Not Siri’s Identity
Google’s Gemini role is real, but it should be understood correctly.
Google announced that Apple developers can now securely call cloud-hosted Gemini models using Apple’s Foundation Models framework and access Gemini in Xcode. That is a meaningful partnership. It gives Apple developers access to stronger cloud-hosted models and gives Google a deeper place inside Apple’s developer ecosystem.
But Gemini’s developer availability does not erase Apple’s own model strategy. Apple’s developer guide says the Foundation Models framework gives developers access to the same on-device model that powers Apple Intelligence and can also work with other language models, including cloud models such as Claude and Gemini, through Apple’s language-model protocol.
That makes the framework a routing layer. Developers can use Apple’s on-device model for private, local, system-integrated tasks. They can use Private Cloud Compute or server-side models when a task needs more power. They can use third-party providers when their app requires a different model capability.
For Siri AI, Apple’s message is that the assistant is built around Apple Intelligence and Apple Foundation Models. Gemini may support parts of the wider AI and developer story, but it is not the brand identity or the entire engine behind Siri.
Foundation Models Framework Becomes the AI Control Layer
The Foundation Models framework may be one of Apple’s most consequential WWDC26 developer announcements because it gives app makers a native path into Apple’s AI stack.
Apple’s developer documentation describes the framework as providing access to Apple’s on-device large language model that powers Apple Intelligence. It can generate text, perform app-specific intelligent tasks, and decide when to call code written by the developer. That last part is especially important because Siri AI and app intelligence are not only about generating words. They are about taking action.
The framework also supports broader model choice. Apple’s WWDC26 developer guide says developers can work with Apple Foundation Models, cloud models like Claude and Gemini, or other providers that conform to the Language Model protocol. This lets Apple keep developers inside its architecture even when they use outside models.
That matters for privacy, security, and consistency. If every developer built AI features through separate third-party SDKs, users would face fragmented permissions and unclear data flows. Apple’s framework gives the company a place to define expectations, tooling, model access, and system behavior.
In other words, Foundation Models is not only an AI feature. It is Apple’s attempt to make AI development feel native to its platforms.
Siri AI Needs App Intents
AFM 3 alone cannot make Siri AI useful. The assistant also needs app-level access, and that is where App Intents becomes essential.
App Intents lets developers expose actions and content from their apps to the system. Apple’s developer materials describe App Intents as a way to connect apps to Apple Intelligence and Siri AI through recognizable structures that make app content discoverable and capabilities available through natural language.
This is the missing piece for a smarter Siri. A model can understand language, but it still needs safe ways to act. App Intents gives Siri AI structured actions instead of forcing the assistant to guess how an app works.
For example, a task app can expose creating a project, adding a deadline, or marking a task complete. A photo app can expose editing actions. A travel app can expose itinerary details. A finance app can expose safe account actions. A smart-home app can expose device controls.
The more developers adopt App Intents, the more Siri AI can do. AFM 3 gives Siri the intelligence layer. App Intents gives Siri the action layer.
Private Cloud Compute Handles the Heavier Requests
Apple’s on-device focus does not mean every request stays on the device. Some AI tasks require more processing power, longer context, or a larger model than iPhone or Mac can efficiently handle. That is where Private Cloud Compute fits.
Private Cloud Compute is Apple’s privacy-focused server architecture for Apple Intelligence requests that need cloud processing. Apple’s message is that users can get more capable AI without turning their personal data into stored cloud records or training material.
This is one of the hard parts of Apple’s strategy. The company wants Siri AI to become more useful, but useful AI often requires personal context. It may need to understand emails, messages, files, photos, calendar events, app content, and what is on the screen. Apple has to process that information while preserving the trust that makes users comfortable with a personal assistant.
AFM 3 does not remove that tension. It gives Apple a stronger local foundation, but cloud support will still matter. The difference is that Apple can choose the route: on device first, Private Cloud Compute when needed, and third-party cloud models only when the app or developer workflow requires them.
That routing is where Apple’s AI privacy story will either succeed or fail.
Dictation and Voice Show Why Local Models Matter
Apple’s AFM 3 Core Advanced announcement specifically mentions expressive voices and higher-accuracy dictation. These may sound like small features compared with broad AI promises, but they are highly practical.
Dictation is one of the most natural uses for on-device AI. It requires speed, accuracy, punctuation, formatting, language understanding, and privacy. Users may dictate messages, notes, emails, reminders, search queries, work drafts, medical details, personal thoughts, or private instructions. Sending all of that unnecessarily to the cloud would weaken Apple’s privacy pitch.
Voice expressiveness is also central to Siri AI. A more natural assistant should not only understand more. It should sound less mechanical, adjust tone better, and respond with more human rhythm. Apple says AFM 3 Core Advanced is natively multimodal, which helps explain why voice and dictation are part of the model story rather than separate side features.
These are not flashy demos. They are everyday interface improvements. A better Siri voice and better dictation can make Apple’s AI feel useful without requiring users to open a chatbot app.
Apple’s AI Strategy Is Hybrid by Design
WWDC26 made clear that Apple’s AI strategy is not one model doing everything. It is a hybrid system.
Apple has on-device models for privacy-sensitive and immediate tasks. It has AFM 3 Core Advanced for its most capable Apple silicon systems. It has Private Cloud Compute for requests that need more power. It has Foundation Models framework for developers. It has App Intents for app actions. It has Gemini and other cloud models available through developer workflows where appropriate.
This is less simple than a single chatbot brand, but it fits Apple’s platform logic. Apple wants the operating system to decide the right tool for the task. The user should not have to think about model selection for every request.
That is also why Siri AI is not just a Gemini wrapper. A wrapper would put another company’s model behind Apple’s interface. Apple is building a system where model routing, app actions, privacy controls, on-device execution, cloud processing, and developer tools all connect through Apple’s platform.
The complexity is hidden behind the assistant. The strategy is in the architecture.
The Privacy Question Is Still Difficult
Apple’s approach is more privacy-conscious than many AI systems, but it is not simple. Siri AI needs more context to become useful, and more context always creates more privacy responsibility.
Users will want to know when Siri is using on-device intelligence, when a request moves to Private Cloud Compute, when a developer app calls Gemini or another model, and what data is being shared. Apple does not need to explain every technical detail in every interaction, but it needs clear user controls and trustworthy defaults.
This is especially important because Apple is inviting developers into the AI layer. If an app uses the Foundation Models framework with Apple’s on-device model, the privacy expectations are different from an app that sends data to a third-party cloud model. Users need transparency, and developers need guidelines that prevent careless data handling.
Apple’s advantage is that it controls the platforms where this will run. Its challenge is that users will judge Siri AI by results. If the assistant is too limited, Apple will look behind. If it is too aggressive with personal context, Apple will look careless. AFM 3 is Apple’s attempt to move faster without giving up the privacy position.
AFM 3 Makes Siri AI an Apple Platform Story
The most important WWDC26 takeaway is that Siri AI is now a platform story, not only an assistant story.
AFM 3 gives Apple its own model family. Foundation Models gives developers a native framework. App Intents connects apps to natural-language actions. Private Cloud Compute gives Apple a way to scale beyond the device. Gemini gives developers optional cloud power. Apple silicon gives on-device models a hardware base.
That combination is much bigger than a Siri redesign. It is Apple building the AI layer of its ecosystem.
The next test will be execution. Users will expect Siri AI to understand more, act faster, respect privacy, work across apps, and feel consistent across iPhone, iPad, Mac, Apple Watch, Vision Pro, and Apple services. Developers will expect tools that are flexible enough for real products, not only demos.
AFM 3 gives Apple a stronger technical answer than simply saying it partnered with Google. The company is using Gemini where it helps, but Siri AI is being positioned as an Apple-native assistant powered by Apple Foundation Models, Apple silicon, and the privacy architecture Apple has spent years building.
That is the difference between borrowing AI and making AI part of the operating system.
