Re: the cloud processing.
It looks like Apple has 3 models. 1 local, 2 cloud.
The cloud model is closed, light weight and the dumbest of the 3. The second model is Apple owned, in the cloud for more compute, and but also something that appears to not learn from the confidential data that people are submitting to it. So this things is probably better than the local model, but still much dumber than GPT or Gemini.
Third model is whatever external LLM you plug into it. OpenAI is the default. Any request to chat GPT throw a consent dialog before any data is sent.