32GB of fast unified memory is enough for Qwen 3.8 27B.
reply- 16GB for the weights at Q4
- 9GB for the full 256K context at Q8
- 7GB spare for overhead and system.
The problem is that these Macs have 32GB of slow unified memory.
Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.
Is this for setup for agentic coding? Why not also run the IDE compiler etc... on the same machine to use those CPU cores as well?
replyKeep in mind that if you want MTP it adds a few gigs. If you use sub-agents it turns already slow generation into even slower generation. Won't be doing any compling (so rust, c and probably go are not avaiable) becase those add memory pressure during compiling.
reply32gb of unified memory is enough enough for system to be used for anything other than LLM generation.
You can, you just need a beefier PC, and it's more annoying in terms of noise and heat vs throwing something on your server closet. Plus you don't need to worry about other software stealing resources and whatnot.
reply> Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.
replyWhat? LLMs are best served from a massive PD disaggregated cluster of B300s connected via NVLink.
If you're running LLMs on a Mac Mini, it's because you want to run local, not because it's the best setup.