DeepSeek · Released May 2025
DeepSeek R1 (0528)
What it takes to run DeepSeek R1 (0528) locally: memory by quantization, the GPUs that handle it, and how fast it runs.
Updated
- Parameters
- 671B (37B active)
- Architecture
- Mixture of experts
- Context
- 128K tokens
- License
- MIT license
- Memory at Q4_K_M
- ~420GB
What do you need to run DeepSeek R1 (0528)?
DeepSeek R1 (0528) needs about 420GB of VRAM at Q4_K_M with 8K context (407GB for the weights plus KV cache and overhead). Because it is a mixture-of-experts model with 37B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.
What to buy
GPUs for DeepSeek R1 (0528)
Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.
Memory
DeepSeek R1 (0528) VRAM requirements by quantization
Lower-bit quantizations shrink the model at some cost to quality. Q4_K_M is the usual sweet spot. Totals include the KV cache for the context length shown plus runtime overhead.
| 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 |
KV cache note: 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.
Performance
How fast DeepSeek R1 (0528) runs
Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.
Overview
About DeepSeek R1 (0528)
DeepSeek R1 was the open reasoning model that shook the industry in January 2025, and the May 2025 (0528) update closed much of the gap to the best closed models of the time. DeepSeek's V4 generation has since superseded it with a 1M-token context, a toggleable thinking mode, and far lower memory per token of context. At 671B parameters it needs about 420GB at Q4_K_M with 8K context, so local use means a 512GB Mac Studio, a large multi-GPU server, or a high-memory EPYC system running the experts from RAM. Aggressive 1-bit to 2-bit dynamic quants roughly halve that at a real quality cost.
Strengths
- +Strong step-by-step reasoning, math, and writing
- +MLA keeps the KV cache compact (about 69 KiB per token)
- +MIT license
Weaknesses
- −Needs about 420GB at Q4_K_M, impractical for most home setups
- −Always thinks, often at great length, which is slow at local speeds
- −Superseded by DeepSeek V4, which is stronger and cheaper to run
- −Text only
Setup
How to run DeepSeek R1 (0528)
$ ollama run deepseek-r1:671b$ llama-server -hf unsloth/DeepSeek-R1-0528-GGUF:Q4_K_M -c 8192 -ngl 99Weights: deepseek-ai/DeepSeek-R1-0528 · GGUF: unsloth/DeepSeek-R1-0528-GGUF
Alternatives
Models like DeepSeek R1 (0528)
DeepSeek
DeepSeek V4 Flash
284B (13B active) · 1.0M ctx
Z.ai
GLM-5.3
744B (40B active) · 1.0M ctx
Alibaba Qwen
Qwen3 235B-A22B (2507)
235B (22B active) · 256K ctx
Z.ai
GLM-5.3-Flash
320B (18B active) · 1.0M ctx
Frequently asked questions
How much VRAM do you need to run DeepSeek R1 (0528)?
About 420GB at Q4_K_M with 8K tokens of context: 406.8GB of weights, 0.6GB of KV cache, and 12.8GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.
Is DeepSeek R1 (0528) free for commercial use?
DeepSeek R1 (0528) is released under the MIT license. That is a permissive license, so commercial use is allowed.
How long a context does DeepSeek R1 (0528) support?
Up to 131,072 tokens (128K). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.
Sources: DeepSeek R1 0528 model card (Hugging Face), Unsloth DeepSeek R1 0528 GGUF files. Capability score basis: Placed just below the gpt-oss-120b anchor from its Artificial Analysis Intelligence Index v3 standing (October 2025).