DeepSeek · 671B parameters · 37B active
DeepSeek-V3 & R1: GPU memory requirements and which GPU to rent
DeepSeek-V3 and its reasoning sibling R1 share one architecture: 671 billion parameters with 37 billion active per token, released in FP8. Multi-head latent attention shrinks the KV cache to about 70 KB per token — the weights, not the context, are what call for an 8-GPU node.
- Weights at FP8671 GB1,342 GB at 16-bit · 369 GB at 4-bit
- FP8, 8k context, 1 sequence700 GBweights + KV cache + runtime
- Suggested here8× H200$10,990 a month · dedicated
Inference
How much GPU memory DeepSeek-V3 & R1 needs to serve
Total memory for one sequence at each context length, and for 16 concurrent sequences at 8k — weights, KV cache at 16-bit and runtime overhead included.
| Weights precision | Weights | 4k context | 32k context | 128k context | 8k × 16 sequences | Suggested here, 8k |
|---|---|---|---|---|---|---|
| 16-bit (BF16 / FP16) | 1,342 GB | 1,397 GB | 1,399 GB | 1,406 GB | 1,406 GB | Multi-node |
| 8-bit (FP8 / INT8) | 671 GB | 700 GB | 702 GB | 709 GB | 709 GB | 8× H200 · $10,990/mo |
| 4-bit (GPTQ, AWQ, Q4) | 369 GB | 386 GB | 388 GB | 395 GB | 395 GB | 2× B300 · $7,580/mo |
GB are 109 bytes. Context lengths are shown for scale where the model supports them; an FP8 KV cache halves the cache part of each figure. As a mixture of experts, every expert stays resident, so memory follows the total parameter count, not the active one. Multi-head latent attention keeps its KV cache unusually small.
Hardware
GPU servers that run DeepSeek-V3 & R1
For FP8 inference with an 8k context and one sequence: the configurations we would suggest, in order — one large card ahead of several small ones.
Fine-tuning DeepSeek-V3 & R1
Memory for a 2k-token sequence and a micro-batch of 4, with gradient checkpointing. Full fine-tuning keeps gradients and Adam states for every parameter; LoRA and QLoRA train small adapters on a frozen base.
| Method | Memory | Suggested here |
|---|---|---|
| QLoRA, 4-bit base | 433 GB | 2× B300 · $7,580/mo |
| LoRA, 16-bit base | 1,445 GB | Multi-node |
| Full fine-tune, mixed precision | 10,806 GB | Multi-node |
Architecture at a glance
Multi-head latent attention keeps its KV cache unusually small. Figures from the model's published configuration; method on the calculator page.
DeepSeek-V3 & R1 hardware: questions answered
How much VRAM does DeepSeek-V3 & R1 need?
For the weights alone: about 1,342 GB at 16-bit, 671 GB at FP8 and 369 GB at 4-bit. Serving it at FP8 with an 8k-token context and one sequence takes about 700 GB in total once the KV cache and the runtime are counted; long contexts and many concurrent users add the KV cache on top — about 0.07 GB per 1,000 tokens at 16-bit.
Can DeepSeek-V3 & R1 run on a single GPU?
Not at FP8: the weights need more memory than any single GPU in our catalogue holds. otherwise the suggested setup is 8× H200 · $10,990/mo.
What is the cheapest GPU setup to run DeepSeek-V3 & R1?
For FP8 inference with an 8k context and one sequence, the configuration we would suggest is 8× H200 · $10,990/mo; at 4-bit, 2× B300 · $7,580/mo. Suggestions favour one large card over several small ones: splitting a model across PCIe cards costs speed.
How much GPU memory do I need to fine-tune DeepSeek-V3 & R1?
With a 2k-token sequence and a micro-batch of 4: about 433 GB for QLoRA on a 4-bit base, 1,445 GB for LoRA on a 16-bit base, and 10,806 GB for a full fine-tune with Adam in mixed precision. Gradient checkpointing is assumed; sharding the optimizer across GPUs spreads the full fine-tune over several cards.
Run DeepSeek-V3 & R1 on 8× H200 for $10,990 a month
Dedicated, single-tenant and live in under 10 minutes — with vLLM, the Hugging Face stack or Ollama preinstalled if you want. Paid in BTC, ETH or USDT, with no KYC.