I have been using Gemini more and more lately at least for public models. I have fallen down the self-hosted rabbit hole and their Gemma models are pretty solid for smaller devices. I run the Gemma3 1B model on my phone and it gives me at least a decent experience for being self-hosted on an iPhone.

I am getting ready to drop $6k on a Mac Studio with 256GB of unified memory that will allow me to run some of the best models in a quantized form, so I am super excited to build and train my own AI.