DeepSeek AI
DeepSeek V4 Pro (671B)
Frontier open-weight MoE with 1M-token context and tool use
Model Summary
Family
DeepSeek
Version
4.0
Parameters
671B (est.)
Parameter counts for closed models are estimates; vendors rarely publish exact sizes.
VRAM Requirements by Quantization
Memory needed to serve DeepSeek V4 Pro (671B) for inference, including a 20% overhead for activations and KV cache.
| Precision | VRAM needed | Smallest single GPU that fits |
|---|---|---|
| INT4 (4-bit) | 374.95 GB | Multi-GPU required |
| INT8 (8-bit) | 749.90 GB | Multi-GPU required |
| FP16 (16-bit) | 1499.80 GB | Multi-GPU required |
| FP32 (32-bit) | 2999.60 GB | Multi-GPU required |
Recommended GPU Configurations
Cheapest on-demand configurations to serve DeepSeek V4 Pro (671B) at 8-bit (750 GB VRAM).
3x AMD Instinct MI325X
768 GB total VRAM · CDNA 3
~$24.00/h
4x AMD Instinct MI300X
768 GB total VRAM · CDNA 3
~$24.00/h
3x AMD Instinct MI355X
864 GB total VRAM · CDNA 4
~$39.00/h
Quick GPU Planning
Use the calculator pre-filled with this exact version to estimate memory, speed, and compute requirements in a few clicks.
Access Pre-filled CalculatorFrequently Asked Questions
How much VRAM do you need to run DeepSeek V4 Pro (671B)?
With an estimated 671B parameters, DeepSeek V4 Pro (671B) needs roughly 750 GB of VRAM in 8-bit (INT8), 375 GB in 4-bit, and 1500 GB in FP16, including a 20% overhead for activations and KV cache.
Which GPUs can run DeepSeek V4 Pro (671B)?
At 8-bit quantization, the most cost-effective option is 3x AMD Instinct MI325X (768 GB combined VRAM, around $24.00/hour on-demand). Higher-end cards like the NVIDIA B200 or AMD MI355X reduce the GPU count needed.
Can DeepSeek V4 Pro (671B) run on a single GPU?
No. Even in 4-bit, DeepSeek V4 Pro (671B) exceeds the memory of any single current GPU, so a multi-GPU cluster is required.
How much does it cost to serve DeepSeek V4 Pro (671B) in the cloud?
Renting 3x Instinct MI325X costs on the order of $24.00/hour, i.e. about $17,520/month running 24/7. Actual prices vary by provider and commitment; spot and reserved capacity can be significantly cheaper.
Deploy on a GPU cloud
Rent 3x Instinct MI325X by the hour instead of buying hardware.
