No question the Mac Studio previously was the flagship Mac regarding performance. Indeed I held off for a while updating my Intel Mac Pro because I intended to replace it with a modern Mac Studio with lots of RAM.
But improvements in processors and especially RAM constraints have changed the ballgame notably. Limiting a Mac Studio to 96 Gb memory with no upgrade path is beyond stupid. I might well reconsider a Mac Studio when higher memory configurations are available. At present I believe the price/performance of available Mac Studio options is a serious fail.
1 Like
Yes, 96 GB is not enough for a lot of people in the research community. It’s driving sales of Nvidia workstations so Apple is losing out.
1 Like
Not just the research community - knowledge workers of all sorts who deal with huge document sets (including attorneys and legal expert witnesses) may benefit from large RAM to run local AI models. I am very strongly considering it myself for this reason.
NVIDIA workstations an more custom Blackwell Unix GPU boxes not only can have more RAM but more importantly they are tremendously expandable to add more RAM and storage in the future.
Such a Unix box is likely to become my “Desktop” whereas my MacBook will be my primary personal workstatin - accessing the Unix box whether I am at home or away.
1 Like
You should ask yourself why you might want to or need to run a local model. Are you concerned about privacy when working with sensitive documents? Are you concerned about cost? Do you want to be able to work offline? Do you just want to tinker with it for fun?
Next: what’s the job to be done, and can you actually do it with a local model? Installing a local model and chatting with it is simple enough, but if you want to tackle the kinds of things you can do with, say, Claude Cowork, you’ll need to set up and manage a local harness, too.
I’m using a Mac Studio with an M2 Max chip with 64 GB of memory. It can comfortably handle a middle tier local model provided nothing else is hogging RAM. Right now I’m using Qwen3.6-27B & Qwen3.6-35B-A3B via LM Studio. They’re good enough for things like summarizing articles or meeting notes, basic document conversion, writing python scripts, working with my DEVONthink databases and the like, but on my Mac at least, they’re slow. If I need to do anything agentic or that needs a lot of reasoning chops, I use Claude Code or Cowork.
3 Likes