How many GPUs do I need to run DeepSeek V3 or R1 (671B)?
SizingUpdated 2026-10-23
DeepSeek V3 and R1 have 671B parameters, about 700 GB in their native FP8, so the usual setup is one 8×H200 node (1,128 GB) or one 8×B200 node; 8×H100 (640 GB) only works with 4-bit quantization or two nodes. At Strata GPU an 8×H200 node costs about $25,170 per month on demand or $21,647 reserved for 24–36 months.
Options
| Setup | GPU memory | Fits |
| 8× H200 | 1,128 GB | FP8 with room for KV cache |
| 8× B200 | ≈ 1,440 GB usable | FP8 with large batches, FP4 option |
| 8× H100 | 640 GB | 4-bit only, or 16× H100 across two nodes for FP8 |
Distilled models
If you need R1-style reasoning without the full model, the distilled 32B and 70B versions run on one or two GPUs. See GPU for a 70B model.
FAQIs a mixture-of-experts model cheaper to serve?
Per token, yes: only about 37B parameters are active per token. But all 671B must still be in GPU memory.
Can I reserve an 8×H200 node for three months?
Yes. Reserved clusters start at one 8-GPU node for three months.