RE: Run even larger AI models locally with LM Studio

avatar
(Edited)

You are viewing a single comment's thread:

Very interesting. I've been looking for an easy to use a large LLM that could use the 64Gb of unified memory on my M1 Max chip.

I don't like interacting with AI in the cloud that collects my data and is not private.

I was surprised to learn that Apple has released a M3 Max with 128Gb unified memory. That would really be powerful and could run huge models.

I'll let you know how it goes.



0
0
0.000
3 comments
avatar

Mac Studio and even Mac Minis are very popular option for LLM due to how unified memory works. Nowhere can you get ~188 VRAM for less than the cost of even a single A100 40G.

0
0
0.000
avatar

I'm getting 23 tokens per second using the 5 bit Mixtal 2.7 model.

0
0
0.000
avatar

macs have a big edge for this.
I would recommend the 4 bit, the 5 bit isn't much better and takes a lot more ram. I'd stick with 4 bit, or something like 8 bit if you can get there.

0
0
0.000