I used to run a lot of local models on my mbp - mainly stt, tts, embeddings and diffusion models - and small LLMs used for utility purposes - but stopped. It saves time in the long run to run those models on target architecture from the get go - which in most cases is nvidia/cuda - rather than test and tweak on metal, and then switch to cuda for prod - and experience weird and subtle differences and regressions. I don't think it makes much sense to develop anything (other than hobby projects for home use) on mlx atm.