DeepSeek-V4-Pro
1,600 billion parameters · MIT license · deepseek-ai/DeepSeek-V4-Pro
The Hugging Face Safetensors file-size counter shows about 862B — that's an artifact of how FP4 weights are packed into files, not the real parameter count. The official model card states 1,600B (1.6 trillion) total parameters, and that is the figure this site uses.
How much GPU memory does this need?
required GB = parameters (billions) × 1 GB, plus 20% working room
1,600 × 1 GB = 1,600 GB
1,600 GB × 1.2 = 1,920.0 GB required
GPU memory is the "workspace" a graphics card uses to hold an AI model while it runs. If a model is bigger than the available memory, it simply will not fit — like trying to fit a large book on a shelf that's too small. More in the wiki →
Minimum setup to run DeepSeek-V4-Pro
- 2,880 GB combined GPU memory (needs ≥ 1,920.0 GB — meets the requirement)
- 28,600 W combined power draw
- $1,000,000 total price
That's as much power as about 23.8 average homes, and about 7.63 electric-car batteries of energy per day of running. An average home draws roughly 1,200 watts around the clock. This compares a build's power draw to that baseline. More in the wiki →
Why more than one machine?
A cluster is just several separate computers wired together and made to work as a team on one job. These are NVIDIA DGX B200 servers, which connect over real datacenter interconnects (NVLink inside each server, InfiniBand between them) built for sharing one model across machines. That's different from just plugging several desktop cards into one PC, which can't share a model this way. More on clusters in the wiki →