RS Advisor · Pick a Model

Pick a model — we'll spec the machine

Open-source models from the arena.ai agent leaderboard (open-source filter). The rule of thumb: GPU memory needed = parameters in billions × 1 GB, plus 20% working room.

Parameters are the size of the model's brain — more parameters usually means a smarter model, but every one of them has to live in GPU memory. GPU memory is the workbench — the whole model must fit here or it doesn't run at all. Regular RAM cannot substitute: it sits on a link far too slow for the GPU, which would spend nearly all its time waiting instead of computing.

DeepSeek-R1

671B parameters · MIT

The memory math, step by step

671B × 1 GB = 671 GB, + 20% working room = 806 GB needed

Minimum setup from our core catalog: 1× NVIDIA DGX B200 = 1,536 GB ✓

Headroom: +730 GB beyond the 806 GB needed.

Why not something smaller? 6× H200 = 846 GB would cover it on paper, but that requires datacenter clustering — see Clusters. The DGX B200 is one box that just works.

Total price: $515,000
Total power: 14,300 W — about 11.9 homes; one day (343 kWh) drains about 3.81 EV batteries

Add to Quote

Kimi K2

1,000B parameters · Modified MIT

The memory math, step by step

1000B × 1 GB = 1000 GB, + 20% working room = 1200 GB needed

Minimum setup from our core catalog: 1× NVIDIA DGX B200 = 1,536 GB ✓

Headroom: +336 GB beyond the 1,200 GB needed.

Why not something smaller? Even 8× H200 = 1,128 GB falls short of the 1,200 GB needed. Nothing smaller than a DGX B200 holds a trillion-parameter brain.

Total price: $515,000
Total power: 14,300 W — about 11.9 homes; one day (343 kWh) drains about 3.81 EV batteries

Add to Quote

Qwen3-235B

235B parameters · Apache 2.0

The memory math, step by step

235B × 1 GB = 235 GB, + 20% working room = 282 GB needed

Minimum setup from our core catalog: 2× NVIDIA H200 SXM = 282 GB ⚠

Honest warning — zero margin: this fits with zero room to spare — 282 GB needed, 282 GB provided. Quantize to 8-bit for breathing room, or step up to 3× H200 (423 GB). Quantization means storing the model's numbers less precisely so it fits in less memory — 8-bit halves it, 4-bit halves it again, with some quality loss.

Why not something smaller? 1× H200 = 141 GB — half of what's needed. 1× B200 = 192 GB — also short. Two H200s pooling memory in one server is the minimum.

Total price: $70,000
Total power: 1,400 W — about 1.17 home; one day (33.6 kWh) drains about 0.37 EV battery

Add to Quote

GLM-4.5

355B parameters · MIT

The memory math, step by step

355B × 1 GB = 355 GB, + 20% working room = 426 GB needed

Minimum setup from our core catalog: 4× NVIDIA H200 SXM = 564 GB ✓

Headroom: +138 GB beyond the 426 GB needed.

Why not something smaller? 3× H200 = 423 GB — just 3 GB short of the 426 GB needed. Close does not count: if the model doesn't fit, it doesn't run. Step to 4.

Total price: $140,000
Total power: 2,800 W — about 2.33 homes; one day (67.2 kWh) drains about 0.75 EV battery

Add to Quote

Llama 3.1 405B

405B parameters · Llama license (open weights)

The memory math, step by step

405B × 1 GB = 405 GB, + 20% working room = 486 GB needed

Minimum setup from our core catalog: 4× NVIDIA H200 SXM = 564 GB ✓

Headroom: +78 GB beyond the 486 GB needed.

Why not something smaller? 3× H200 = 423 GB, but 486 GB is needed. 4× H200 in one GPU server gives 564 GB.

Total price: $140,000
Total power: 2,800 W — about 2.33 homes; one day (67.2 kWh) drains about 0.75 EV battery

Add to Quote

A cluster is several machines wired together so they can act like one bigger machine.

The honest part: real datacenter clustering works because NVLink and InfiniBand let GPUs share memory at enormous speed. Eight desktop cards in a closet do NOT make a datacenter — the links between them are so slow that a big model split across them runs painfully slowly. Desktop cards are for models that fit on ONE card.