I am very excited by this, but I am a bit dampened that the maximum memory available is 128GB. I was really hoping for 256GB, which would allow me to run frontier models locally. I think with 128GB it's still feasible to use this with something like Qwen3-Coder-Next and MiniMax-M2.5, but things like Kimi-K2.5 will require significant quantization to fit and model performance will really suffer.
I'm really wanting to build proper local-first AI workflows at home, and I think Apple has an opportunity to make that possible in a way other companies aren't really focused on, but we need significantly larger memory capabilities to do it, which I know is tough in the current memory market but should be available for a cost.
I spent the last day deep diving on what I can do with MLX with local models. I still feel limited, because you have to use quantized models, but I think it's enough to do /something/, so I went ahead and bit the bullet and pre-ordered just now. I am driven a little bit by concern about ongoing memory market pressures over the next 1-3 years, and thinking it's a bit now or never.
I'm really wanting to build proper local-first AI workflows at home, and I think Apple has an opportunity to make that possible in a way other companies aren't really focused on, but we need significantly larger memory capabilities to do it, which I know is tough in the current memory market but should be available for a cost.