Apple reveals new AI model family built with Google and Nvidia

Apple has unveiled its third generation of Apple Foundation Models (AFM), a family of five AI models that power a redesigned Siri and a range of new intelligent features across its operating systems. The announcement came at the company’s annual Worldwide Developers Conference on Monday. For its most capable model, Apple worked with both Google and Nvidia, as multiple sources including Kif Leswing reporting for CNBC and MacRumors detail.

The new model family includes two on-device models and three server-based models. All server-based models run on Apple’s Private Cloud Compute infrastructure, which the company says ensures that user data is never stored or shared with anyone, including Apple itself.

Five models, two tiers

• AFM 3 Core is the updated version of Apple’s 3-billion-parameter on-device model.
• AFM 3 Core Advanced is a new 20-billion-parameter on-device model with a sparse architecture, activating only 1 to 4 billion parameters at a time. It is natively multimodal and enables features like expressive voices and higher-accuracy dictation without any cloud requests.
• AFM 3 Cloud is the standard server-side model, optimized for speed and efficiency.
• ADM 3 Cloud (Image) handles image generation and editing, powering tools like Image Playground and spatial reframing in Photos.
• AFM 3 Cloud Pro is the most capable model in the family, designed for complex reasoning and agentic tasks such as multi-step actions across apps.

AFM 3 Cloud Pro is the model where Apple’s partnerships with Google and Nvidia come into direct play. To run it, Apple extended its Private Cloud Compute infrastructure to Nvidia GPUs hosted in Google’s cloud. Apple software VP Sebastien Marineau-Mes explained that a recent Nvidia technology called “confidential compute” allowed the chips to be configured so they cannot read the contents of Apple’s servers, meeting Apple’s privacy requirements.

Apple AI executive Amar Subramanya said AFM 3 Cloud Pro is comparable in quality to Google’s Gemini frontier models. However, Apple executives were quick to clarify what Google’s role actually was. Craig Federighi, Apple’s SVP of Software Engineering, stated: “The amount of the Google Assistant we use is none.” Apple uses none of the Gemini models deployed to Google’s own customers, none of Google’s client-side code, and no Google Search infrastructure.

According to Subramanya, the four Apple Silicon models were “custom built for Apple Silicon, trained using proprietary data with reinforcement learning and refined using outputs from Gemini frontier models.” Google’s contribution was therefore a form of knowledge distillation, not a direct integration of its products.

Apple also described a System Orchestrator, a piece of software that routes every query to the right model based on complexity and how much personal data it requires. Federighi called it “key to the privacy architecture of our entire system.” For queries about current events, the system draws on Apple’s own World Knowledge Service, which the company says it has been building for several years.

Apple published internal evaluation results to support its quality claims. In human preference tests, AFM 3 Cloud was preferred over last year’s server model on 64.7 percent of prompts, compared to 8.7 percent for the older model. The on-device AFM 3 Core was preferred over its predecessor on 45.6 percent of prompts, versus 23.3 percent for the previous version. For its new expressive voices feature, AFM 3 Core Advanced scored 4.15 on a standard 5-point audio quality scale, compared to 3.87 for the existing system.

Apple said it does not use users’ private data or interactions to train its models and that all Private Cloud Compute infrastructure, including the extended Nvidia capacity in Google’s cloud, can be independently verified by third-party researchers.

The company plans to publish a full technical report with additional benchmarks later this summer.

Sources

• Introducing the Third Generation of Apple’s Foundation Models – Apple Machine Learning Research
• Apple partnering with Google and Nvidia for most advanced AI model – CNBC
• Apple’s New AI Models Contain ‘None’ of Google’s Gemini Assistant – MacRumors