Apple launches third generation of AFM models: local, cloud, and on Google servers with Nvidia

What happened?

During the WWDC26 keynote, Apple announced the third generation of its Apple Foundation Models (AFM). These are five large language models (LLMs) designed for different environments: some run entirely on-device, others on Apple's Private Cloud Compute, and for the first time, one is hosted on Google Cloud servers using Nvidia accelerators. According to 9to5Mac, this move allows Apple to offer generative AI capabilities without compromising privacy in the most sensitive use cases, while relying on third-party infrastructure for computationally intensive tasks.

Why is this important?

Historically, Apple has prioritized on-device processing to ensure privacy. However, the most advanced language models require computing power that no mobile chip can offer today. By splitting the workload between local, its own cloud, and Google's cloud with Nvidia, Apple achieves a balance: it answers simple queries without sending data externally, scales with its own cloud for moderate tasks, and delegates the heaviest ones to Google, leveraging Nvidia's optimization for inference. This could redefine the industry standard, where until now companies like Samsung and Google rely exclusively on their own data centers or partnerships with cloud providers.

Market implications

Apple's decision has several implications. First, it reinforces Google Cloud's position as an infrastructure provider for direct competitors in AI, which could create tensions with other clients. Second, Nvidia consolidates its dominance in AI acceleration, now also within the Apple ecosystem. Third, iOS and macOS app developers will be able to integrate more powerful AI capabilities without worrying about privacy, as Apple maintains its anonymization layer. Finally, users will see enhanced features like improved Siri, image generation, and smart summaries with reduced response times.

What readers should know

Not all third-generation AFM models will be available immediately. Apple has confirmed that the on-device and Private Cloud Compute models will arrive with iOS 26 and macOS 26, while the model hosted on Google Cloud will be progressively enabled for specific tasks such as complex multimedia content generation. Users will be able to control which data is sent to the cloud and will have the option to disable external processing entirely. Additionally, Apple assures that no user data is stored on Google servers, and all communication is end-to-end encrypted.

"Apple has found a formula that combines privacy, performance, and flexibility. By using Google servers with Nvidia, it avoids investing in massive proprietary infrastructure and accelerates its time to market. It's a win-win for the end user," commented an analyst from TheVortiq.

Historical perspective

Compared to the launch of the first AFMs in 2024, which only operated on-device, this third generation represents a qualitative leap. In 2025, Apple introduced Private Cloud Compute, but it was still limited to its own servers with Apple Silicon chips. Now, by opening up to Google and Nvidia, Apple acknowledges that there are tasks its hardware cannot handle efficiently. This pragmatism contrasts with its traditional secrecy and could signal a new era of strategic AI collaborations.

Frequently asked questions

• What specific models have been announced? Apple has not detailed the names, but it is known there are five models: two on-device (one lightweight for basic tasks and one larger for advanced functions), two on Private Cloud Compute (for moderate processing), and one on Google Cloud with Nvidia (for heavy tasks).
• How does this affect privacy? Apple states that data never leaves the device for local models; in its own cloud, it is processed with anonymization; and on Google Cloud, end-to-end encryption is applied and no data is stored.
• When will they be available? With iOS 26 and macOS 26 in fall 2026, although the Google Cloud model may arrive in later phases.