Apple has unveiled the technology powering the next generation of Apple Intelligence, and one detail stands out above everything else: the company’s newest AI foundation models were developed in collaboration with Google.
Announced at WWDC 2026, Apple’s third-generation Apple Foundation Models will serve as the backbone for Siri AI, Image Playground, advanced photo editing tools, smarter system-wide intelligence, and many of the AI features coming to iPhone, iPad, Mac, Apple Watch, and Apple Vision Pro later this year.
While most of the attention at WWDC focused on the redesigned Siri AI experience, Apple quietly revealed one of the biggest shifts in its AI strategy yet. The company has moved to a new architecture centered around five foundation models that span both on-device processing and cloud-based intelligence, with Google playing a key role in their development.
The new lineup includes AFM 3 Core, a 3-billion-parameter on-device model designed for everyday Apple Intelligence features, and AFM 3 Core Advanced, a far more powerful 20-billion-parameter model built to unlock advanced capabilities on Apple’s most capable hardware.
Apple says AFM 3 Core Advanced introduces a new architecture that allows much larger AI models to run efficiently on consumer devices. Instead of keeping the entire model active in memory, it selectively loads only the components needed for a specific task. This approach helps improve performance while reducing memory requirements and power consumption.
Beyond on-device AI, Apple is also introducing three new cloud-based models that run through Private Cloud Compute. AFM 3 Cloud handles general AI requests, while AFM 3 Cloud Pro is designed for more demanding reasoning tasks and advanced tool usage. Apple also introduced ADM 3 Cloud, a dedicated image model that powers photorealistic image generation, image editing, Genmoji, and new creative tools across the Apple ecosystem.
Perhaps the most surprising detail is that Apple extended its Private Cloud Compute infrastructure to NVIDIA GPUs running in Google Cloud for AFM 3 Cloud Pro. According to Apple, the company has maintained the same privacy guarantees that prevent user data from being stored or accessed, even while leveraging more powerful cloud hardware.
These new foundation models are what make the revamped Siri AI possible. Apple says the upgraded assistant can engage in natural conversations, understand personal context, reason across multiple sources of information, and take actions across apps. The company also highlighted stronger multimodal capabilities, allowing Siri AI to understand images, screenshots, documents, and other visual content.
Image generation and editing are receiving major upgrades as well. ADM 3 Cloud powers the new Image Playground experience, advanced photo editing features, Spatial Reframing, and photorealistic image creation. Apple showcased examples featuring realistic people, detailed environments, and complex lighting effects that represent a significant leap beyond previous Apple Intelligence image tools.
Apple also shared early evaluation results showing substantial improvements over its previous generation of models. According to the company, human evaluators consistently preferred responses generated by the new AFM 3 Core and AFM 3 Cloud models. Apple also reported major gains in image understanding, dictation accuracy, text-to-speech quality, and instruction-following performance.
Despite the deeper AI integration, Apple continues to position privacy as a key differentiator. The company reiterated that it does not use private user data or personal interactions to train its foundation models. Requests are handled either directly on-device or through Private Cloud Compute, where Apple says user information remains inaccessible to both Apple and third parties.
The announcement marks one of the most significant AI updates in Apple’s history. While Siri AI may be the feature users interact with most, the real story is the technology underneath it. Apple’s new foundation models, developed alongside Google and optimized for Apple silicon, will power nearly every major Apple Intelligence feature introduced across iOS 27, iPadOS 27, macOS 27, watchOS 27, and visionOS 27.
