This guide provides the definitive hardware specifications required to run DeepSeek R1 locally, from the lightweight distilled versions to the full-scale 671B parameter model.
Running LLMs locally is primarily limited by VRAM (Video RAM). Below is the minimum VRAM required for various quantization levels (4-bit vs 8-bit).
| Model Version | Parameters | Min VRAM (4-bit) | Recommended GPU |
|---|---|---|---|
| R1-Distill-Qwen | 7B | 6 GB | RTX 3060 / 4060 |
| R1-Distill-Llama | 70B | 40 GB | 2x RTX 3090/4090 |
| DeepSeek R1 (Full) | 671B | 320 GB+ | 8x A100 / H100 |
The easiest way to test DeepSeek R1 on your machine is via Ollama.
# To run the 7B version (Ideal for 8GB VRAM cards)
ollama run deepseek-r1:7b
# To run the 32B version (Requires ~20GB VRAM)
ollama run deepseek-r1:32b
💡 Hardware Optimization Tips Unified Memory: Apple Silicon (M2/M3/M4 Max) is highly efficient for DeepSeek because the GPU can access the entire system RAM.
Quantization: Use GGUF or EXL2 formats to reduce the model size by 50-70% with minimal loss in "intelligence".
Flash Attention: Ensure your environment supports Flash Attention 2 to speed up inference by up to 2x.
📖 Deep Dive & Performance Benchmarks For a full breakdown of token-per-second (TPS) benchmarks across different NVIDIA and Apple hardware, read our detailed analysis:
👉 Read the full guide: DeepSeek R1 Hardware Guide
What you will learn in the full article: Multi-GPU Setup: How to link two RTX 4090s for the 70B model.
Mac vs. PC: Detailed ROI analysis for local LLM development.
Power Consumption: Managing the heat and electricity of a local AI server.
📄 License
Distributed under the MIT License. See LICENSE for more information.