DeepSeek V4 Flash
Confirmed on DeepSeek's official Hugging Face model card: 284B total parameters (13B "active" per response -- MoE models like this only wake up part of themselves for any given answer, but the whole 284B still has to sit in memory).
Minimum GPU memory needed
The 20% extra isn't optional padding — running a model also needs working memory for the conversation itself and temporary calculations, on top of the model's own weights.
What actually runs it
Four workstation cards clear the memory bar for far less than any datacenter server -- the cheapest way to run this model on our shelf. See the honest note on clustering below before you buy four of anything.
This is a cluster
A cluster just means multiple GPUs or machines working together on one job instead of one machine doing it alone. Real datacenter clustering connects GPUs with NVIDIA's NVLink — extremely fast, direct links that let them share memory and act like one giant GPU. Simply owning several separate cards side by side isn't the same thing: without NVLink between them, they talk over a much slower general-purpose connection and can't pool memory as seamlessly. It works, but it's not true datacenter-grade clustering.
≈ 2.0 average homes running continuously (an average home draws about 1,200 watts around the clock). Over one full day that's 57.6 kWh — about 0.64× a typical electric car battery (roughly 90 kWh).
Alternative: 1x HGX B200 Server
Massive headroom, pre-clustered with NVLink out of the box, but far pricier than it needs to be for this model alone.