With reports about Apple Intelligence V1 not having met quality expectations, I wondered if the (adapter-based architecture was to blame: depending on the use-case, a different adapter would be plugged into the base model. A newer report confirms, however, that the new “Apple Intelligence V2” will also use the adapter-based approach:

For specialized use cases that require teaching the ~3B model entirely new skills, we also provide a Python toolkit for training rank 32 adapters.

That adapters continue to provide for particular use-cases is described in the “Responsible AI” section:

To design our mitigation steps for multilingual use, we began with multilingual post-training alignment at the foundational model level, then extended to feature-specific adapters that integrate safety alignment data.

The core base model continues to be around 3B in size, but V2 adds long-context attention mechanisms (RoPE, NoPE) and Vision Encoder(s) and adapter. (For the server-sider, a novel parallel Mixture-of-Experts architecture (PT-MoE) is used.)

So in conclusion: no, Apple did not abandon adapters - quite the opposite, it seems.