Running DeepSeek-R1 Locally: Quantization & VRAM Math Guide
DeepSeek-R1 has revolutionized open-weights reasoning. This hands-on guide details hardware VRAM formulas, GGUF quantization tradeoffs (Q4_K_M vs Q8_0), and one-command local deployments with Ollama.
1. VRAM Requirements & Hardware Formula
Calculate required VRAM using the standard formula:
Required VRAM (GB) = (Params in Billions * Bits per Weight / 8) * 1.20 (KV-Cache Overhead)
- DeepSeek-R1 14B (Q4_K_M): ~9 GB VRAM (Runs on RTX 4070 / 3080).
- DeepSeek-R1 32B (Q4_K_M): ~20 GB VRAM (Runs on RTX 4090 / 3090).
- DeepSeek-R1 70B (Q4_K_M): ~43 GB VRAM (Requires 2x RTX 3090 or Mac Studio M3 Max 64GB).
2. Quickstart Ollama Commands
# Run 14B distilled reasoning model
ollama run deepseek-r1:14b
# Run 32B model for software engineering tasks
ollama run deepseek-r1:32b