[bigyabai]

It might run the smaller flash version, but 96gb is not enough for the trillion-parameter model.

The M3 Ultra's GPU is a bit on the weak side for large-scale inference, so you'll be waiting on token prefill for most coding/agent workflows.

[namegulf]

They have a 512gb ram option but pricey.

Have you tried any other models with this M3 Ultra?

[bigyabai]

The 512gb model would have to use a lobotomized quant like q_2 or q_1, and you would still be waiting 3-5 minutes to process context lengths in the 32,000-64,000 token range.

Apple's GPUs are just not very fast for inference. I'd stick to the smaller 7b-18b parameter range or MOE models like Qwen if you want a usable inference speed.

[namegulf]

Looks like that's a good idea for now. Yeah 3-5 mins is not practical use.

Any thoughts on M5?

They may be soon releasing a M5 model with mac studio/mini.

[namegulf]

NVIDIA DGX Spark a good option?

$4,699.00

But looks like we may need a NVIDIA AI Enterprise - DGX Spark License