Searching for the exact DeepSeek R1 local hardware requirements often surfaces conflicting numbers because “DeepSeek-R1” refers to two distinct model families: the full 671-billion-parameter Mixture-of-Experts (MoE) flagship (37B active parameters per token) and the six dense DeepSeek-R1-Distill models (`1.5B`, `7B`, `8B`, `14B`, `32B`, and `70B`) fine-tuned on Qwen and Llama architectures. While our DeepSeek R1 Architecture Guide breaks down Multi-Head Latent Attention (MLA) and GRPO training math, this practical deployment guide provides the exact GPU VRAM, system RAM, disk space, and copy-paste Ollama + Open WebUI commands for every model size.
Interactive DeepSeek-R1 Model Size & Ollama Command Generator
Select any DeepSeek-R1 variant (`1.5B` through `671B`) and deployment runtime to inspect exact VRAM/RAM requirements and copy the launch command.
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama run deepseek-r1:14b
Complete DeepSeek-R1 Local Hardware Requirements Table (1.5B to 671B)
When Ollama runs ollama run deepseek-r1 without a tag, it defaults to deepseek-r1:7b (the 4.7 GB Qwen-7B distilled model). Use this complete hardware sizing table to pick the highest-parameter variant that fits 100% inside your physical GPU VRAM or Apple Silicon unified memory:
| Ollama Tag | Base Architecture | Q4_K_M Disk Size | VRAM Needed (8K–16K Ctx) | Minimum Target Hardware |
|---|---|---|---|---|
deepseek-r1:1.5b | Qwen-2.5-1.5B Dense | 1.1 GB | ~2.2 GB | Any 4GB GPU, integrated laptop iGPU, or CPU-only |
deepseek-r1:7b | Qwen-2.5-7B Dense | 4.7 GB | ~6.2 GB | 8GB VRAM GPU (RTX 3060/4060, RX 7600) or M1/M2 8GB |
deepseek-r1:8b | Llama-3.1-8B Dense | 4.9 GB | ~6.5 GB | 8GB VRAM GPU or Apple Silicon 16GB |
deepseek-r1:14b | Qwen-2.5-14B Dense | 9.0 GB | ~11.2 GB | 12GB–16GB VRAM GPU (RTX 3060 12GB, 4070, 4060 Ti 16GB) |
deepseek-r1:32b | Qwen-2.5-32B Dense | 19.8 GB | ~22.4 GB | 24GB VRAM GPU (RTX 3090, RTX 4090, RX 7900 XTX) |
deepseek-r1:70b | Llama-3.3-70B Dense | 42.5 GB | ~46.5 GB | 2× 24GB GPUs (48GB) or Apple Silicon 64GB+ Unified |
deepseek-r1:671b | DeepSeekMoE (37B Active) | 404 GB (Q4) / 131 GB (1.58-bit) | 160 GB – 430 GB | Mac Studio 192GB/512GB or Multi-GPU + DDR5 Server (`ktransformers`) |
Step-by-Step: Running DeepSeek R1 Locally with Ollama + Open WebUI
To get a ChatGPT-style interface that renders DeepSeek-R1’s collapsible <think>...</think> reasoning blocks cleanly on your local machine without sending a single token to external servers:
# Step 1: Pull your target DeepSeek-R1 model in Ollama (e.g., 14B for 12GB-16GB GPUs)
ollama pull deepseek-r1:14b
# Step 2: Create a custom Modelfile with DeepSeek's recommended 0.6 temperature (prevents repetition loops)
cat << 'EOF' > Modelfile.deepseek
FROM deepseek-r1:14b
PARAMETER temperature 0.6
PARAMETER top_p 0.95
PARAMETER num_ctx 16384
EOF
ollama create deepseek-r1-14b-tuned -f Modelfile.deepseek
# Step 3: Launch Open WebUI bound to localhost (127.0.0.1:3000)
docker run -d -p 127.0.0.1:3000:8080
--add-host=host.docker.internal:host-gateway
-v open-webui:/app/backend/data
--name open-webui --restart always
ghcr.io/open-webui/open-webui:mainNeed to calculate exact memory footprints for custom context lengths (`32K` or `64K`) or compare other coding models? Use our 2026 Local LLM VRAM Calculator and our Best Local LLMs for 8GB, 12GB & 16GB VRAM Guide.
Written and prepared by Kshitij Gupta.
