Apple and Google Partner on New AI Foundation Models
At its annual Worldwide Developers Conference (WWDC) this week, Apple announced a major overhaul of its artificial intelligence platform, revealing that its new Apple Foundation Models (AFMs) were co-developed using Google’s Gemini technology. This collaboration marks a significant shift in Apple’s AI strategy, powering a conversational reconstruction of Siri and a brand-new suite of features under the Apple Intelligence umbrella.
Key Takeaways
- Shared Training, Proprietary Execution: Apple’s new AI models are trained and distilled using Google’s Gemini architecture, but they run entirely on Apple’s proprietary code without active third-party software or Google code in the execution loop.
- Efficient Local Architecture: The new AI framework includes AFM Core Advanced, a 20-billion parameter on-device model that preserves battery life by dynamically activating only 1 to 4 billion parameters for any given request.
- Strict Privacy in the Cloud: The most powerful model, AFM Cloud Pro, runs on NVIDIA hardware inside Google Cloud, but it uses Apple’s Private Cloud Compute security to ensure that no user data is shared with Google, saved on external servers, or visible to Apple.
- Flexible Model Swapping: Starting in iOS 27 and macOS 27, a new LanguageModel protocol allows third-party developers to natively replace Apple’s local models with alternative platforms such as Gemini or Anthropic’s Claude.
- Native Developer Tools: Google’s Gemini is now natively integrated into Apple’s Xcode development environment, enabling programmers to write, test, and debug code with agentic AI assistance without leaving their primary workspace.
While the partnership represents a deep technical alliance, Apple executives clarified that the new models are not simply a licensed version of Google’s software. Instead, Apple utilized Gemini’s architecture and capabilities to train and distill its own proprietary models, which are highly optimized for Apple Silicon.
Craig Federighi, Apple’s Senior Vice President of Software Engineering, emphasized that when users interact with the base Apple Intelligence features, they are not running Google code on their devices or interacting with an active Gemini agent. The final software stack remains entirely under Apple’s control.
The resulting AI architecture consists of five distinct models designed for various tasks. For local processing, Apple introduced AFM Core, a standard model for day-to-day on-device tasks.
Beside it is AFM Core Advanced, a 20-billion parameter local model designed for complex, on-device multimodal tasks such as voice and image processing. To preserve battery life and performance, the Advanced model uses a sparse activation system, only utilizing between 1 and 4 billion parameters for any given request.
For tasks requiring more computational power, Apple introduced its server-side suite. This includes AFM Cloud for general-purpose processing, AFM Cloud Image for generating and editing graphics, and AFM Cloud Pro for highly complex reasoning tasks.
Notably, the Cloud Pro model represents the peak of the Apple-Google partnership, achieving performance comparable to Google’s frontier models.
This hybrid setup raises questions about data privacy, especially since the Cloud Pro model runs on NVIDIA graphics processors located inside Google’s cloud infrastructure. Apple has addressed these concerns by extending its Private Cloud Compute architecture to these external servers.
By utilizing confidential computing technologies, Apple guarantees that user data is processed securely and is never shared with Google, stored on external servers, or accessible to Apple itself.
In addition to system-level upgrades, Apple is making it easier for developers to work with alternative AI platforms. Starting with iOS 27 and macOS 27, a new LanguageModel protocol allows third-party developers to easily swap Apple’s local models with external models, such as Gemini or Anthropic’s Claude, using standard development interfaces.
Furthermore, Apple and Google have integrated Gemini directly into Xcode, allowing developers to debug code, write software, and build apps with AI assistance natively inside Apple’s development environment.
About the Author: Julio Caesar
