Best home rig for DeepSeek R1 671B

Will it fit my GPU? Best value GPU per euro →

Real, measured generation speeds for DeepSeek R1 671B on home hardware, fastest first. Different quants/engines aren't directly comparable.

Build	Quant	Engine	tok/s	Cost
Mac Studio M3 Ultra 512GB (Geerling test)		llama.cpp	19.89		source
Mac Studio M3 Ultra, 512 GB	Q4	MLX	18.43		source ⚠ unchecked
EPYC + single RTX 3090 (hybrid)	Q2_K	llama.cpp CUDA	8.97		source ⚠ unchecked
$2K EPYC 64-core, 512 GB DDR4, CPU-only	Q4	llama.cpp (bare metal Ubuntu 24)	4.31	$2,000 (≈€1,784)	source
AmpereOne A192-32X 512GB (CPU only)		llama.cpp	4.18		source

Rent before you buy: try this model on a cloud GPU for a few dollars first: Runpod (referral link; sign up with Google to get the new-user bonus)

Memory math: how much VRAM / RAM does DeepSeek R1 671B need?

DeepSeek R1 671B has 671B parameters, about 37B active per token (source: Ollama library: 671B total, 37B active (V3 base)). Weight size = parameters × bits-per-weight ÷ 8, before context (KV cache) and runtime overhead.

Format	Bits/weight	Weights	Smallest common memory size that holds the weights alone
Q8_0	8.50	713.0 GB	1024 GB
Q6_K	6.56	550.5 GB	1024 GB
Q5_K_M	5.70	478.4 GB	512 GB
Q4_K_M	4.89	410.5 GB	512 GB
Q3_K_M	4.00	335.2 GB	512 GB
Q2_K	3.16	265.0 GB	512 GB

Bits-per-weight for GGUF quants are llama.cpp's own figures (llama.cpp quantize README). "As measured" rows apply the source's stated bits-per-weight to the parameter count; real files differ slightly because some layers stay at higher precision. Try other combinations in the VRAM calculator.

Why it can still run on smaller GPUs: it is a mixture-of-experts model, so only about 37B of its 671B parameters are used for each generated token. Runtimes such as llama.cpp (--n-cpu-moe), ExLlamaV3 or Project Maya can keep most expert weights in system RAM and only the shared layers and hot experts in VRAM. Then you need VRAM + system RAM ≥ weights, and speed is limited mainly by RAM bandwidth.

Measured on setups with less GPU memory than the Q4_K_M weights (so part of the model lived in system RAM or was streamed from disk):

RTX 3090 (24 GB): 8.97 tok/s (Q2_K)

FAQ

How much VRAM does DeepSeek R1 671B need?

The weights are about 410.5 GB at Q4_K_M (671B params × 4.89 bits ÷ 8, before context/KV cache). Add memory for context (KV cache) and the runtime. The smallest-memory setup with a measured result here, fully in memory, is Apple M3 Ultra (512 GB) at 19.89 tok/s.

What is the fastest measured hardware for DeepSeek R1 671B?

Apple M3 Ultra: 19.89 tok/s generation (llama.cpp), from https://github.com/geerlingguy/ai-benchmarks/issues/42.

Which GPUs have measured results for DeepSeek R1 671B?

Apple M3 Ultra, RTX 3090.

Can it run on… Apple M3 Ultra · RTX 3090