DeepSeek R1 Hardware Requirements: What You Need to Run It Locally (2026)
On this page
Short answer: the full DeepSeek R1 model (671B parameters) needs about 400 GB of memory even when compressed, which puts it out of reach of normal PCs. What most people run at home are the distilled R1 models (1.5B to 70B), which are much smaller models trained to imitate R1's reasoning. A gaming PC with an 8 GB graphics card handles the 7B/8B versions well; 24 GB of VRAM or a 32 GB+ Mac runs the 32B version.
DeepSeek R1 sizes and the hardware they need
Sizes below are the default 4-bit (Q4_K_M) downloads from Ollama. Allow a bit of extra memory on top for the context (the conversation).
| Model (Ollama tag) | Download size | GPU (NVIDIA/AMD) | Mac (unified memory) | Good for |
|---|---|---|---|---|
| 1.5B (deepseek-r1:1.5b) | 1.1 GB | Any, or CPU only | 8 GB | Testing, very old PCs |
| 7B (deepseek-r1:7b) | 4.7 GB | 6–8 GB VRAM (RTX 3060/4060) | 16 GB | Maths and simple reasoning |
| 8B (deepseek-r1:8b) | 5.2 GB | 8 GB VRAM | 16 GB | Best small all-rounder |
| 14B (deepseek-r1:14b) | 9.0 GB | 12–16 GB VRAM (RTX 3060 12GB, 4070) | 16–24 GB | Noticeably smarter, still fast |
| 32B (deepseek-r1:32b) | 20 GB | 24 GB VRAM (RTX 3090/4090) or 32 GB (RTX 5090) | 32–48 GB | Strongest single-GPU option |
| 70B (deepseek-r1:70b) | 43 GB | 48 GB VRAM (2 × 24 GB GPUs) | 64 GB+ | Close to big-model quality |
| 671B full model (deepseek-r1:671b) | 404 GB | Multi-GPU server (e.g. 8 × 80 GB) | 512 GB (Mac Studio, top config) | The real R1 |
The 8B tag in Ollama now points to DeepSeek-R1-0528-Qwen3-8B, an updated May 2025 distill, and the 671B tag to the updated R1-0528. For a detailed comparison of the two small models, see DeepSeek R1 7B vs 8B.
What matters most: memory, not processor speed
- VRAM (graphics memory) is the key number. If the whole model fits in VRAM, responses are fast. If it doesn't, part of it runs on the CPU and speed drops sharply.
- System RAM should be at least as large as the model when running on CPU, and 16–32 GB is a sensible minimum for a GPU setup.
- Apple silicon Macs share memory between CPU and GPU, so a Mac with 32–64 GB of unified memory can run models that would need an expensive graphics card on a PC.
- Storage: use an SSD with enough free space; model files load much faster from NVMe. If space runs low, see where Ollama stores models.
Can you run DeepSeek R1 without a GPU?
Yes, but slowly. The 1.5B and 7B/8B models run on a modern CPU with 16 GB of RAM at a few words per second, which is usable for short questions. Larger models on CPU only are very slow, and the reasoning models "think" for a long time before answering, which makes slow hardware feel even slower.
Running the full 671B model at home
- Standard 4-bit (404 GB): needs around 400–450 GB of fast memory, such as a Mac Studio with 512 GB of unified memory or a multi-GPU server.
- Extreme compression (Unsloth dynamic quants): the smallest 1.66-bit version of R1-0528 is about 162 GB. Unsloth recommends at least 180 GB of combined RAM + VRAM for around 5 tokens per second; with 64 GB of RAM and no GPU expect about 1 token per second.
For most people, the full model is cheaper and faster through DeepSeek's website, app or API, or through cloud providers that host it.
Recommended setups by budget
| Budget | Setup | What it runs well |
|---|---|---|
| What you already have | Any PC/laptop with 16 GB RAM | 1.5B, 7B/8B (slowly on CPU) |
| Entry | Gaming PC with RTX 3060 12 GB / RTX 4060 Ti 16 GB, or MacBook with 16–24 GB | 8B, 14B |
| Enthusiast | RTX 3090/4090 (24 GB), RTX 5090 (32 GB), or Mac with 48–64 GB | 32B comfortably; 70B on a 64 GB Mac |
| Workstation | 2 × 24 GB GPUs or Mac with 128 GB+ | 70B at good speed |
| Server | Mac Studio 512 GB or multi-GPU server | Full 671B |
How to run it
- Install Ollama or LM Studio (see LM Studio vs Ollama).
- In Ollama, run for example
ollama run deepseek-r1:8b. Pick the largest size that fits your VRAM from the table. - For better answers, DeepSeek and Unsloth recommend a temperature around 0.6 (top_p 0.95).
Should you still run R1 in 2026?
DeepSeek has released newer models since R1: the V3.1/V3.2 line combined chat and reasoning, and the DeepSeek V4 family (April 2026) is open-weight too. But V4 is far too large for home PCs: even the smaller V4-Flash has about 284B total parameters, and V4-Pro about 1.6 trillion. For local use, the R1 distills are still among the most practical reasoning models, alongside other small open models. For the newest DeepSeek models, use the official app or API. Our DeepSeek R1 vs V3 guide explains how the reasoning and chat models differ.
FAQ
How much RAM do I need for DeepSeek R1?
16 GB of system RAM for the 7B/8B models, 32 GB for 14B–32B on a GPU, and about 400 GB of fast memory for the full 671B model at 4-bit.
Can an 8 GB GPU run DeepSeek R1?
Yes, the 7B and 8B distills (about 5 GB) fit in 8 GB of VRAM.
Can I run DeepSeek R1 on a laptop?
Yes. Gaming laptops with 8 GB+ VRAM run the 8B model well, and Apple silicon MacBooks with 16 GB or more run 7B–14B.
Is the distilled model the same as DeepSeek R1?
No. Distills are smaller Qwen and Llama models trained on R1's outputs. They reason well for their size but are clearly weaker than the full 671B R1.