Apple used its WWDC26 keynote to unveil the third generation of its Foundation Models, a five-model lineup that now stretches from the device to Google Cloud. The company said AFM 3 Cloud Pro runs on NVIDIA GPUs hosted in Google Cloud, marking the first time its Private Cloud Compute architecture has been extended beyond Apple infrastructure.
The announcement lands now because Apple device owners and developers are already searching for what changes in the company’s AI stack, and the answer is bigger than a new model name. AFM 3 Core and AFM 3 Code Advanced are built for on-device use, while AFM Cloud, ADM 3 Cloud (Image) and AFM 3 Cloud Pro handle server-based tasks. Apple says the new system keeps its security and privacy protections intact, which matters because the company’s original cloud AI setup kept processing inside Apple data centers on Apple silicon servers.
Apple first laid out its foundation models in 2024, when the lineup included an on-device language model with roughly 3 billion parameters and a larger server-based language model running through Private Cloud Compute on Apple silicon servers. This new generation is broader and more ambitious. Apple said all five models start from a common foundation before branching into different architectures and use cases, and it added multimodal capabilities including audio, image understanding, long-context reasoning and high-quality visual generation.
The biggest technical shift is also the most delicate one. Apple is presenting AFM 3 Cloud Pro as part of a privacy-preserving system even though it runs on external infrastructure in Google Cloud. That puts the company in a new place: still insisting on the same security model, but relying on third-party hardware for part of the AI workload. Apple and Google also worked on capabilities that go beyond a traditional confidential computing deployment, making the partnership more than a simple hosting deal.
Inside the new lineup, AFM 3 Core Advanced is the standout on-device model. Apple said it packs 20 billion parameters into a model that activates up to 4 billion parameters at a time depending on the prompt, using a sparse architecture. That design points to where Apple is headed next: more powerful AI features that can run locally when possible, and cloud support when the task is too large. What remains unclear is how much of Apple’s future AI traffic will ultimately depend on Google-hosted infrastructure.
For users, the relevance is immediate. Apple is tying the next phase of Apple Intelligence-style features to a model family built for both privacy and scale, the same kind of selling point that has already shaped interest in iphone 17 pro coverage and broader Siri AI plans. For developers, the five-model system signals a more flexible foundation to build on. The unanswered question is not whether Apple has a new AI stack. It is how far Apple will let that stack leave its own servers from here.
