Hardware guide · October 2026
What GPU do you need to run DeepSeek R1 (0528)?
DeepSeek R1 (0528) is a 671B (37B active) mixture-of-experts model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run DeepSeek R1 (0528) at Q4_K_M with 8K tokens of context you need about 420GB of VRAM (406.8GB of weights plus KV cache and overhead). On a Mac or other unified-memory system, the M3 Ultra 512GB runs it at ~14 tokens/sec.
- VRAM at Q4_K_M
- ~420GB
- GPUs that run it
- 2
- Parameters
- 671B
- CPU only (est.)
- ~2.0 tok/s
Our recommendations
The best GPUs for DeepSeek R1 (0528)
Mac / unified memory
M3 Ultra 512GB512GB unified · $9,499 MSRP
- Quant
- Q3_K_M
- Speed (est.)
- ~14 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
DeepSeek R1 (0528) VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 252 GB | 260 GB | 262 GB |
| Q3_K_M | Usable | 327 GB | 338 GB | 340 GB |
| Q4_K_M | RecommendedDefault | 407 GB | 420 GB | 422 GB |
| Q5_K_M | Very good | 478 GB | 494 GB | 495 GB |
| Q6_K | Near-lossless | 554 GB | 571 GB | 573 GB |
| Q8_0 | Practically lossless | 713 GB | 736 GB | 737 GB |
| BF16 | Full precision | 1,342 GB | 1,383 GB | 1,385 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, DeepSeek R1 (0528) needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.3 GB | 419.9 GB | 419.7 GB |
| 8K tokens | 0.6 GB | 420.2 GB | 419.9 GB |
| 16K tokens | 1.2 GB | 420.7 GB | 420.2 GB |
| 32K tokens | 2.3 GB | 421.9 GB | 420.7 GB |
| 64K tokens | 4.6 GB | 424.2 GB | 421.9 GB |
| 128K tokens | 9.2 GB | 428.8 GB | 424.2 GB |
Multi-head latent attention caches one compressed 576-value latent per layer (512 for keys and values plus 64 for the RoPE key): 61 layers x 576 x 2 bytes = 70,272 bytes per token, about 9.2GB at 128K. Runtimes without MLA support expand this to full per-head keys and values, roughly 70 times larger.
Compatibility
Every GPU that runs DeepSeek R1 (0528)
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| M5 Ultra 512GB | 512GB unified | Q3_K_M | ~21 tok/s | Fits, tight | — |
| M3 Ultra 512GB | 512GB unified | Q3_K_M | ~14 tok/s | Fits, tight | $9,499 |
Tips
Getting the most out of DeepSeek R1 (0528)
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Offload experts, not layers. For mixture-of-experts models, llama.cpp's
--n-cpu-moe(or-ottensor overrides) keeps attention on the GPU and moves expert weights to system RAM. That keeps speed usable when the model does not fit in VRAM. - Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the DeepSeek R1 (0528) overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run DeepSeek R1 (0528)?
About 420GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 736GB, and at Q3_K_M about 338GB with some quality loss. Longer contexts add 0.07GB per 1K tokens with a 16-bit KV cache.
Can I run DeepSeek R1 (0528) on a 24GB GPU?
No. DeepSeek R1 (0528) needs about 420GB at Q4_K_M, which does not fit on a 24GB card like the RTX 3090 or RTX 4090, even with CPU offload on a 64GB system.
Can I run DeepSeek R1 (0528) on a Mac?
Yes. With 512GB of unified memory or more (for example the M3 Ultra 512GB), macOS can give the GPU enough memory to run it at Q3_K_M, at about ~14 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run DeepSeek R1 (0528) on CPU only, without a GPU?
Technically yes on a server with enough RAM, but at about 420GB it needs a multi-channel server platform and will be slow on a desktop.