Apple Details New Foundation Models, Clarifies Complete Independence from Google's Gemini
Key Takeaways
- ▸Apple confirmed it uses none of Google's Gemini models, infrastructure, or Google Search in its foundation models—using Gemini only for distillation-based model refinement
- ▸New AFM family includes two on-device models (AFM Core and multimodal AFM Core Advanced) and three server-side models, all trained on proprietary data with reinforcement learning
- ▸Apple partnered with Nvidia and Google to extend Private Cloud Compute to Nvidia GPUs in Google's cloud, with security guarantees preventing Google from accessing Apple's data
Summary
Apple executives have detailed the architecture of the company's new Apple Foundation Models (AFM) family and clarified exactly how Google's technology factored into their development. Addressing press concerns, Craig Federighi, Apple's SVP of Software Engineering, stated unequivocally that Apple uses "none" of Google's Gemini models, client-side code, or search infrastructure. The company's third-generation AFM family spans two on-device models (AFM Core and AFM Core Advanced) and three server-side models (AFM Cloud, AFM Cloud Image, and AFM Cloud Pro), all custom-built for Apple Silicon and trained using proprietary data with reinforcement learning.
While Google's involvement is limited to distillation-based refinement using Gemini frontier models, Apple's infrastructure approach reflects deeper partnerships. To run AFM Cloud Pro—the most capable model designed for agentic tool use—Apple extended its Private Cloud Compute to Nvidia GPUs hosted in Google's cloud, leveraging Nvidia's "ambiguous confidential compute" technology to ensure Google cannot access Apple's server contents. The broader system features a System Orchestrator that intelligently routes queries to appropriate models based on complexity and personal context, drawing on Apple's own World Knowledge Service for current events.
- System Orchestrator intelligently routes queries to models based on complexity and context, with all Private Cloud Compute infrastructure independently verifiable by third-party researchers
Editorial Opinion
Apple's explicit clarification about its independence from Gemini signals a strategic pivot to owning its entire AI stack end-to-end. While the use of Gemini outputs for distillation is a pragmatic acknowledgment of frontier model quality, Apple's emphasis on proprietary data, custom silicon optimization, and privacy-preserving infrastructure positioning sets a clear competitive narrative. This move reflects growing industry tension over AI model dependency and Apple's commitment to differentiating on privacy—a critical advantage as users become increasingly concerned about where their data flows.
