Hardware to run DeepSeek V4-Pro 1.6T (MoE)

Apr 2026 release. 1.6 T / 49 B active, 1 M-token context. 67.9% Terminal-Bench — top open-weights agent, matches closed frontier on most agentic tasks. Q2 ~430 GB.

Quantization Q2 Q2_K Q4 Q4_K_M Q5 Q5_K_M Q8 Q8_0

Availability ⚡ Only plug & play

Cheapest

2× Mac Studio M3 Ultra 512 GB cluster (TB5 / MLX)

Apple · two desktops, Thunderbolt 5 RDMA

$28,400

tokens / secQ4

235B-MoE 29 t/s

671B-MoE 9.0 t/s

1T-MoE 7.0 t/s

Memory1024 GB · 960 usable

Bandwidth819 GB/s

Idle / Active24 W / 440 W

Sticker$28,400

Why: Lowest sticker that still fits DeepSeek V4-Pro 1.6T (MoE) ($28k USD).

Apple RefurbB&H Photo (used)eBay (used)

📺 Reviews on YouTube

▶ Apple didn't have to go this hard… (4× M3 Ultra TB5 cluster) ▶ Mac Studio CLUSTER vs M3 Ultra

Fastest

DGX B200 — 8× B200 server (1.44 TB HBM3e)

NVIDIA · 10U DGX server

$475,000

tokens / secQ4

235B-MoE 165 t/s

671B-MoE 105 t/s

1T-MoE 90 t/s

Memory1440 GB · 1404 usable

Bandwidth8000 GB/s

Idle / Active900 W / 10200 W

Sticker$475,000

Why: Highest measured tg/s — 59 t/s on DeepSeek V4-Pro 1.6T (MoE)-class models at Q4.

NVIDIA DGX B200SuperMicro HGX B200Lambda Labs

📺 Reviews on YouTube

▶ Inside a 1.44 TB HBM3e NVIDIA HGX B200 AI Server from ASRock Rack

All-rounder

12× RTX Pro 6000 Blackwell rack (1152 GB)

NVIDIA · 8U server rack (multi-node, 1-2 chassis)

$118,000

tokens / secQ4

235B-MoE 260 t/s

671B-MoE 95 t/s

1T-MoE 55 t/s

Memory1152 GB · 1116 usable

Bandwidth1792 GB/s

Idle / Active340 W / 7400 W

Sticker$118,000

Why: Top quartile across speed, value, memory headroom, and efficiency — the "buy this if unsure" pick.

AmazonCoreWeaveSuperMicroSHI enterprise

Best value

8× DGX Spark cluster (1024 GB unified, CUDA)

NVIDIA · rack of 8 desktops, 200 GbE fabric

$43,500

tokens / secQ4

235B-MoE 42 t/s

671B-MoE 16 t/s

1T-MoE 13 t/s

Memory1024 GB · 976 usable

Bandwidth273 GB/s

Idle / Active220 W / 1840 W

Sticker$43,500

Why: Best $/tg-per-second — ~$5,124 per t/s.

Amazon (8× DGX Spark)Micro CenterNVIDIA MarketplaceServeTheHome guide

📺 Reviews on YouTube

▶ NVIDIA didn't want me to do this — 8× DGX Spark 1 TB VRAM cluster (AZisk) ▶ I built an 8x NVIDIA GB10 cluster for massive Local AI

Best CUDA

DGX H200 — 8× H200 server (1.13 TB HBM3e)

NVIDIA · 8U DGX / HGX server rack

$380,000

tokens / secQ4

235B-MoE 155 t/s

671B-MoE 100 t/s

1T-MoE 85 t/s

Memory1128 GB · 1100 usable

Bandwidth4800 GB/s

Idle / Active700 W / 6500 W

Sticker$380,000

Why: Strongest CUDA-only software stack among fitting builds.

NVIDIA DGX H200SuperMicro HGXLambda Labs

📺 Reviews on YouTube

▶ Inside the SUPER NVIDIA H200 Server From Supermicro (8U HGX H200)

No plug-and-play build fits at Q4_K_M

Only used / DIY / homelab-cluster rigs fit DeepSeek V4-Pro 1.6T (MoE) at this quant. Turn off "Only plug & play" to see them.

Last updated 2026-07-11