A few notes from things I learned in WWDC25 labs relevant to what you wrote:
- Engineers said we can expect the models to be updated with each new version of the OS. They also said we should have time during beta of those OS to test for regressions, and especially have time to update our adapters.
However during beta 1, something contradictory happened: users of 26 beta 1 (all platforms) got an update directly in Apple Intelligence settings.
Which means that supposedly they are also capable of updating it without pushing a new OS version.
- Compared to MLX, performance should be better indeed. Foundation Models run on the neural engine, (and I assume it can also mobilize CPU and GPU), whereas MLX models can only run currently on CPU + GPU. That should give a nice performance boost.
Excellent article Natasha, thanks!
A few notes from things I learned in WWDC25 labs relevant to what you wrote:
- Engineers said we can expect the models to be updated with each new version of the OS. They also said we should have time during beta of those OS to test for regressions, and especially have time to update our adapters.
However during beta 1, something contradictory happened: users of 26 beta 1 (all platforms) got an update directly in Apple Intelligence settings.
Which means that supposedly they are also capable of updating it without pushing a new OS version.
- Compared to MLX, performance should be better indeed. Foundation Models run on the neural engine, (and I assume it can also mobilize CPU and GPU), whereas MLX models can only run currently on CPU + GPU. That should give a nice performance boost.
Thanks - great info to add!