hmm it is fun to play around with local "ai" models but 32gb memory is a quite tight fit while lighter models at least in my testing didn't return sufficiently good results (not even sure about qwen3.8 27B-Turbo right now but at least it supports thinking)
then the even more pressing issue (context size limitation problems aside) is tok/s - the framework 13 is simply too slow. the m2pro mac mini would be faster but i wouldnt be able to free as much of the 32gig of memory i am afraid. the next best thing would be an m4pro macbook but this is limited to only 24gig of memory 😩

...and the tower with only rx570 cards .... well i think that can be ignored