EDGE
Inference keeps moving to where the work happens
Platform-level language models now run on the device with structured output and tool use. That makes local execution a real application boundary, while larger or more demanding work still belongs elsewhere.
What changed
Apple’s Foundation Models framework gives applications direct access to an on-device language model for tasks such as summarisation, extraction, structured generation, and tool calling. Apple says the on-device model works offline and keeps its inputs and outputs on the device.
That is a concrete shift in the application boundary. Local inference is no longer limited to a separate model server maintained by an AI team. It can be a capability exposed by the operating system beside the files, messages, sensors, and user actions that give the task its context.
Why the edge keeps winning some work
Moving the model closer can remove a network round trip, reduce the data that must leave the device, and keep a feature available when the connection is poor or absent. Those advantages matter most for frequent, bounded tasks with local context: classify this item, extract these fields, summarise this document, or prepare a structured action for the user to approve.
The boundary is not absolute. Apple itself directs applications that need more reasoning capability or context toward Private Cloud Compute or another server provider. A sovereign design therefore needs routing, not a slogan: decide what must remain local, what can leave, and what happens when the preferred model is unavailable.
What we are watching
- Whether applications disclose the local, private-cloud, and external paths clearly enough for people to understand them.
- How developers test quality across device generations and model versions.
- Which workloads stay useful offline rather than merely technically possible.