284B parameters. No GPU.

DeepSeek V4 Flash has 284 billion parameters. Someone got it running on a PC with 64 GB of RAM and no graphics card. Two ideas make that possible.

Problem one: it doesn't fit

As released, the model is about 160 GB. A PC with 64 GB of RAM can't hold that.

So the author squeezed it. Every parameter is a number, and the release stores most of them in 4 or 8 bits. The community build stores them in 1.71 bits on average: most experts drop to about 1.5 bits, while some sensitive parts stay at 4 to 5 bits. That brings the whole model down to about 57 GB, which just about fits, with a little room to spare.

Squeezing this hard can cost some quality. The author's own test scores it well, but people who run heavily compressed models often report a noticeable step down below about 4 bits. Test it on your own tasks before relying on it.

Problem two: too much maths per token

Even if it fits, a CPU can't push all 284 billion numbers through for every token it writes.

It doesn't have to. V4 Flash is a mixture of experts. In each layer there are 256 small "expert" networks, and a router picks a few of them for each token: 6 routed experts plus 1 shared expert. Everything else in that layer sits idle for that token.

	Parameters
Total	~284B
Used per token	~13B

So the work per token is closer to a 13B model than a 284B one. On a Ryzen 9 laptop CPU, the author reports around ten tokens a second.

The stadium of 256 pillars in the video is a simplified illustration: the real model picks experts in every one of its layers, and experts don't specialise by topic like "maths" or "code".

Why RAM matters more than a GPU here

The catch with a mixture of experts: the router can pick any expert for the next token, so all of them have to sit in memory, even though only a few run at a time. That makes memory the hard limit and compute the easy part.

For this one model, on this one machine, the answer to "what do I need?" was more RAM, not a graphics card.

The speed and memory numbers are the author's own measurements and haven't been independently verified.