Apple Foundation Models Use Gemini for Training, Not Architecture

Apple confirmed this week that its third-generation Apple Foundation Models were trained using outputs from Gemini as a distillation source, not built on top of Google’s products. The distinction matters more than the headline suggests.

Craig Federighi was unusually precise in his language during the post-keynote briefing, listing exactly what Apple does not use: Google’s deployed Gemini models, Google’s client-side code, and Google Search as a knowledge backbone. What Google actually contributed was model distillation, meaning Apple used Gemini frontier outputs to refine its own models during training, a common technique that transfers capability without transferring architecture or data pipelines. The original Apple Intelligence rollout drew criticism for leaning too heavily on third-party AI, so the specificity here reads partly as reputation management.

The more interesting disclosure is the infrastructure arrangement for AFM Cloud Pro, Apple’s most capable server model. To run it, Apple extended its Private Cloud Compute environment onto Nvidia GPUs hosted inside Google’s cloud, relying on a technology called “ambiguous confidential compute” to prevent the hardware from reading Apple’s server contents. That means Apple’s most powerful AI model runs on Google’s physical infrastructure, which is a stranger arrangement than the Gemini training story.

The Model Family

The full AFM lineup breaks down like this:

• AFM Core: on-device, dense architecture
• AFM Core Advanced: on-device, sparse and natively multimodal, handles features like expressive voices locally
• AFM Cloud: latency-optimized for standard Private Cloud Compute requests
• AFM Cloud Image: handles image generation and editing, including spatial reframing
• AFM Cloud Pro: agentic reasoning, runs on Nvidia GPUs in Google’s cloud

The System Orchestrator, which routes queries to the appropriate model, is where Apple’s privacy architecture actually lives. It pulls from a Spotlight Semantic Index, on-screen context, and an in-house World Knowledge Service built over several years, keeping personal context local while pushing only complexity-appropriate requests to the cloud. Users setting up Apple Intelligence for the first time are interacting with this routing layer every time Siri responds.

Apple says all Private Cloud Compute infrastructure, including the Google-hosted Nvidia capacity, remains open to independent verification by third-party researchers. Whether that assurance extends meaningfully to hardware Apple does not own is a question the company has not fully answered.