Your computer can run it, but you have to layer the model across CPU and GPU memory. Your bottleneck will be PCIe speed, which probably won't be a huge issue for a 4080 on smaller quants.
How are you going to fit 671 billion parameters into 64 + 16 gigs of RAM? That's less than one bit per parameter.
Deepseek v3 is an MoE model - not every parameter is activated at the same time.
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.
What would you suggest instead? I also thinking about.
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.