How much VRAM do I need to run DeepSeek V3?
DeepSeek V3 is a 671B parameter mixture-of-experts model, so it needs roughly 1.3TB of VRAM at 16-bit and around 350GB to 400GB even at 4-bit. That means a multi-GPU node in every case: a cluster of 80GB H100 or A100 cards, or a smaller set of 140GB-class cards like the H200 or MI300X. Only about 37B parameters are active per token, so it runs faster than its size suggests, but the whole model still has to fit in memory. The table below lists the highest-memory GPUs we track for building such a node.
| GPU | VRAM | $/hr | Where | |
|---|---|---|---|---|
| MI300X | 192 GB | $2.19 | RunPod secure cloud | Rent → |
| B200 | 180 GB | $3.50 | Vultr on-demand | Rent → |
| H200 | 143 GB | $3.62 | Massed Compute on-demand | Rent → |
Because DeepSeek V3 is a mixture-of-experts model, its full 671B parameters all live in VRAM even though only a fraction activate for any given token. At about 2GB per billion parameters at 16-bit, that is roughly 1.3TB, so full-precision serving spans a large multi-GPU node, commonly eight or more 80GB cards or several MI300X at 192GB each.
Quantization is what makes it practical to self-host. At 4-bit the footprint falls to around 350GB to 400GB, which a node of high-memory cards can hold, with the usual modest tradeoff in output quality. The MoE design keeps inference efficient because only the active experts do work per token, so throughput can be strong once the model fits.
The binding constraint is total VRAM across the node, not the speed of any single card. The table below lists the largest-memory GPUs in our index, sorted by price, as the building blocks. For how a full multi-card instance is priced, see the node pricing explainer.
Related questions
- What GPU do I need to run Llama 405B?
- How does multi-GPU node pricing work?
- What GPU do I need for a 70B model?
Numbers on this page come from today's verified snapshot. Full table on the homepage; method in the methodology.