← Interactive LLM VRAM Calculator — try your own context length & quantization

How much VRAM to run DeepSeek-R1 671B? (FP16/Q8/Q6/Q4 requirements)

Architecture: MoE with 671B total / ~37B active parameters, MLA compressed KV cache (61 layers). Figures are approximations for weights + an 8k-token FP16 KV cache + ~5 GB runtime overhead, based on typical published GGUF sizes. Actual usage varies by runtime and settings.

QuantizationWeightsTotal VRAM (8k context)Fits on (example)
FP16
full precision (16-bit)
~1342 GB~1348 GBmulti-node GPU cluster (~1.4 TB)
Q8_0
8-bit
~711 GB~717 GB≥10×A100/H100 80 GB or 4×MI300X 192 GB
Q6_K
~6.6-bit
~550 GB~556 GB8×A100/H100 80 GB (640 GB)
Q5_K_M
~5.7-bit
~476 GB~482 GB8×A100 80 GB or Mac Studio 512 GB
Q4_K_M
~4.8-bit
~404 GB~410 GB6×A100/H100 80 GB or Mac Studio 512 GB

How much VRAM do I need to run DeepSeek-R1 671B?

You need roughly ~410 GB of GPU or unified memory to run DeepSeek-R1 671B at Q4_K_M — for example 6×A100/H100 80 GB or a Mac Studio with 512 GB of unified memory. At its native FP8 precision the weights alone are ~713 GB. The full DeepSeek-R1 is a 671B-parameter MoE model: only ~37B parameters are active per token, but ALL 671B must be loaded in memory. Its MLA attention keeps the KV cache small (well under 1 GB at 8k), but the weights alone put it firmly in multi-GPU-server or 512 GB-unified-memory territory. Note the original release is FP8 (~713 GB), so the FP16 row is a theoretical conversion.

What affects these numbers?

Three things: the quantization (bytes per parameter), the KV cache (grows linearly with context length — the table assumes 8k tokens), and runtime overhead. Longer contexts need more memory; quantizing the KV cache (e.g. q8_0 cache in llama.cpp) roughly halves that part. Use the interactive calculator to plug in your exact context length and quantization.