Filed the bug with 'Feedback Assistant'.

Fired the bug.

I profiled the app in 'Instruments', found that it was using the CPU to inference, so it is much slower than the original model.