Apple introduced the third generation of its own models, a family of five built jointly with Google, covering execution both on the device and on servers.
Developing models in partnership with the company that also supplies the product's flagship model is an arrangement answering a practical limitation: training a frontier model requires infrastructure few organisations worldwide have assembled, and building from scratch would take years.
The split between on-device and server execution is the technically interesting part. An on-device model responds fast, works without a network and sends no data outside, resolving much of the privacy concern. In exchange, it fits less capability.
That split is the same one any operation faces when designing a system with AI: what needs a large model and what can be solved by a small one running close to where the data is.
The period brought evidence that the locally executable slice grew, with small models designed to work without cloud and demonstrations of models running even on cheap microcontrollers.