4 ms·

How are you going to fit 671 billion parameters into 64 + 16 gigs of RAM? That's less than one bit per parameter.

• talldayo 2y ago
  Deepseek v3 is an MoE model - not every parameter is activated at the same time.
  • smallerize 2y ago
    But different experts will probably be activated between every token. You will have to load data from disk to generate the next token. That's painfully slow.
    • theGeatZhopa 2y ago
      What would you suggest instead? I also thinking about.
      • smallerize 2y ago
        Not to be flippant, but paying for access to an API is going to be way more cost-effective than buying sufficient hardware to run this locally. You might try waiting to see if someone distills this model down to a smaller one, or sparsifies it by deleting a bunch of parameters and fine-tuning, or comes up with a 1.5-bit version. But you're going to need 2x RAM to run this one even very optimistically. 4x would give better quality outputs. And it will still be fairly slow unless you throw more GPU hardware at it.