Apple silicon has a pretty decent on-board ML subsystem, you can get LLMs to output a respectable number of tokens per second off of it if you have the memory for them. I’m honesty shocked that they haven’t built a little LLM to power Siri