DeepSeek R1 671B VRAM Requirements
DeepSeek R1 671B needs about 642 GB of VRAM at FP8 (official) with an 8K context: 641 GB for weights, 0.5 GB for the KV cache and 0.9 GB of runtime overhead. At Q4_K_M it needs 380 GB. At Q8_0 it needs 665 GB and at full FP16 precision 1251 GB.
642 GB VRAM
DeepSeek R1 671B at FP8 (official) with 8K context and the KV cache in FP16.
- Weights 641 GB
- KV cache 0.5 GB
- Runtime overhead 0.9 GB
Mixture of experts: all 671B parameters need memory, but only 37B are read per token, so it runs at the speed of a much smaller model.
Which GPUs can run it
- RTX 3060 12GBToo small
- RTX 4060 Ti 16GBToo small
- RTX 3090 24GBToo small
- Radeon RX 7900 XTX 24GBToo small
- RTX 4090 24GBToo small
- RTX 5090 32GBToo small
- RTX 6000 Ada 48GBToo small
- A100 80GBToo small
- H100 80GB SXMToo small
- Ryzen AI Max+ 395 128GBToo small
- Apple M4 Max 128GBToo small
- Apple M5 Max 128GBToo small
- RTX PRO 6000 Blackwell 96GBToo small
- NVIDIA DGX Spark 128GBToo small
- Apple M3 Ultra 512GBToo small
DeepSeek R1 671B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| FP8 (official) | 8.21 | 641 GB | 642 GB | 644 GB | Multi-GPU / server |
| Q4_K_M | 4.85 | 379 GB | 380 GB | 382 GB | Multi-GPU / server |
| Q5_K_M | 5.69 | 444 GB | 446 GB | 447 GB | Multi-GPU / server |
| Q6_K | 6.59 | 515 GB | 516 GB | 518 GB | Multi-GPU / server |
| Q8_0 | 8.50 | 664 GB | 665 GB | 667 GB | Multi-GPU / server |
| FP16 / BF16 | 16.00 | 1250 GB | 1251 GB | 1253 GB | Multi-GPU / server |
| Q3_K_M | 3.91 | 305 GB | 307 GB | 308 GB | Multi-GPU / server |
| Q2_K | 3.00 | 234 GB | 236 GB | 237 GB | Multi-GPU / server |
VRAM by context length
At FP8 (official). This model uses Multi-head Latent Attention: each layer caches one compressed 576-value vector per token instead of full keys and values. Runtimes without MLA support use many times more memory.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.3 GB | 0.1 GB | 642 GB | 642 GB |
| 8K tokens | 0.5 GB | 0.3 GB | 642 GB | 642 GB |
| 32K tokens | 2.1 GB | 1.1 GB | 644 GB | 643 GB |
| 128K tokens | 8.6 GB | 4.6 GB | 650 GB | 646 GB |
GPUs that can run DeepSeek R1 671B
No single listed GPU or Mac holds it at FP8 (official).
About DeepSeek R1 671B
The full reasoning model, a 671B mixture-of-experts with 37B active parameters. It uses Multi-head Latent Attention, which compresses the KV cache to 576 values per token per layer, about 69 KB per token in FP16. Runtimes without MLA support store full keys and values and need many times more memory for context. The official checkpoint is FP8, about 641 GiB.
| Spec | Value |
|---|---|
| Parameters | 671B total, 37B active (MoE) |
| Layers | 61 |
| Hidden size | 7168 |
| Attention heads / KV heads | 128 / MLA latent (576) |
| Head dimension | 128 |
| Vocabulary | 129,280 |
| Max context (used here) | 128K tokens |
| KV cache per token (FP16) | 69 KB |
| Released | 2025-01 |
| License | MIT |
DeepSeek R1 671B VRAM: common questions
Can I run the full DeepSeek R1 at home?
Only with very large unified memory or a server: about 380 GB at Q4_K_M. Community dynamic 1.58-bit and 2-bit quants bring it to roughly 130 to 230 GB at a quality cost.
Are the DeepSeek R1 distill models the same as R1?
No. The distills are Qwen and Llama models fine-tuned on R1's outputs. They are far smaller and have their own pages here.
How much VRAM does DeepSeek R1 671B need?
About 642 GB at FP8 (official) with an 8K context, 665 GB at Q8_0 and 1251 GB at FP16. Longer contexts add 0.07 GB per 1,000 tokens with an FP16 cache.
Can DeepSeek R1 671B run on a single GPU?
Not at FP8 (official). It needs about 642 GB, which takes a multi-GPU server or a very large unified-memory system.
Does DeepSeek R1 671B need VRAM for all 671B parameters?
Yes. All experts must be loaded because any token can use any expert. Only 37B parameters are read per token, so it generates text about as fast as a 37B dense model.