How much VRAM does DeepSeek-R1 need?

DeepSeek-R1 has 671B parameters, so its weights alone take about 1340 GB of VRAM at FP16 — or roughly 335 GB quantized to INT4. Real serving adds KV cache on top, which scales with your context length and concurrent requests. Size the exact figure for your workload below.

Uses Multi-head Latent Attention (MLA), which compresses KV cache far below standard GQA math. This tool's KV estimate uses the generic GQA formula as a conservative upper bound. Reasoning models also produce long chain-of-thought output tokens — size avgOutputTokens generously.