Apple’s new Foundation Models explained: on-device AI, cloud AI, and everything in between
Apple’s third-gen Foundation Models split tasks between on-device and cloud systems, including a new model running on Google Cloud. The lineup is built around privacy, speed, and more capable AI features.
Intelligence analysis by GPT-5.4 Mini
Apple’s WWDC26 AI push centers on five Foundation Models: two on-device, two Apple-cloud models, and one image model, plus a more powerful cloud model running on NVIDIA GPUs in Google Cloud. Apple says the system keeps its privacy protections while expanding into multimodal, agentic, and higher-reasoning tasks.
Apple made a family of AI helpers. Some live inside the device like a small notebook, and some live in the cloud like a giant library. The smartest jobs go to the bigger helpers, but Apple still wants the private parts kept safe.
Analysis
What Apple announced
Apple’s third-generation Foundation Models (AFM) line up as five separate models: two on-device models, two Apple server-side models, and one image model. The on-device side includes AFM 3 Core and AFM 3 Code Advanced, while the cloud side includes AFM Cloud, ADM 3 Cloud (Image), and AFM 3 Cloud Pro.
On-device vs. cloud
The key split is where the work happens. Apple says most of the models are built to run on Apple silicon devices, but AFM 3 Cloud Pro runs on NVIDIA GPUs in Google Cloud. That is notable because Apple extended its Private Cloud Compute architecture to third-party infrastructure for the first time while saying it still preserves Apple’s security and privacy protections.
Why AFM 3 Core Advanced stands out
AFM 3 Core Advanced is Apple’s most powerful on-device model. Apple describes it as a 20-billion-parameter model, but it does not keep all of that active at once. Instead, it uses a sparse architecture that activates only part of the model depending on the request. The article says this means roughly 1 to 4 billion parameters are active at a time, which helps make a larger model practical on-device.
Why AFM 3 Cloud Pro matters
AFM 3 Cloud Pro is Apple’s most capable server-based model. Apple says it is meant for demanding use cases such as agentic tool use and complex reasoning. The article also notes that Apple’s security blog frames the Google Cloud deployment as a layered privacy system, with attestation, hardware transparency, and isolated components meant to reduce risk.
Bigger picture
Apple says all five models started from a shared base before specializing for tasks like audio, image understanding, long-context reasoning, and visual generation. That suggests Apple is trying to build one AI stack that can stretch from a watch-sized device to a cloud-heavy reasoning task without losing the privacy story that defines its platform.
Key points
- Apple’s third-generation Foundation Models include five models split across on-device and server-based use cases.
- AFM 3 Core Advanced is a 20-billion-parameter on-device model that uses sparse activation to stay practical.
- AFM 3 Cloud Pro runs on NVIDIA GPUs in Google Cloud and is meant for the most demanding tasks.
- Apple says it extended Private Cloud Compute to third-party infrastructure while keeping security and privacy protections.
- The models were built from a shared foundation and then specialized for audio, image understanding, reasoning, and generation.
If Apple’s approach works, users could get faster AI for simple tasks on-device and stronger help for harder jobs in the cloud. The mix of sparse on-device models and cloud systems could let Apple add more capable features without giving up its privacy-first message.
The setup also raises execution risk: Apple now has to make a split system feel seamless across devices, Apple cloud, and Google Cloud. If privacy guarantees, performance, or model quality fall short, the complexity could weaken the very experience Apple is trying to improve.
