llm’s are random text generators. each one is a different architecture, different training data, or different training recipe.
first off: you probably dont need to fill up your harddrive with every good model in existance. often a url link is sufficient.(unless the owner deletes it. but quants are often still there)
optional second questions: how many links do you have to decent models(excluding low value experiments) how many would you estimate that you have?
optional third question: how much is your divirsity of model usage? assuming you had the compute, how many tested models would you utilize for a single query across different models? average. max.
fourth question: do you still rely on inference compute for non-locally hosted models?
Local llms would be big boot to your privacy if nothing else. You dont know what cloud llm providers do with your info, how long they store your data, or who they sell it to. The only HIPPA compliant model I know of is Claude series through api key access specifically. With local llms you dont have to worry about those things.
Also consider they changing and tweaking models all the time behind the scenes what works great for you now may change next update. Once you download a local llm its yours no chance of changes happening to important therapy not.
Chain of thought models are really good for use cases where you want quality answers over speedy ones.