Model advisor

Pick the model you want to run; we show you the minimum you must buy to run it, with the memory math in the open. All three are big open-source models from the world agent leaderboard, MIT-licensed — free to download and run on your own hardware.

The memory rule, in one line

GPU memory needed = parameters in billions × 1 GB, plus 20% working room.

Why this rule works

A model's size is its parameter count — the "B" in a name like DeepSeek V4 Flash 284B means 284 billion learned numbers. Stored at 8 bits each, every billion parameters takes about 1 GB. The extra 20% is working room: the scratch space the model uses to remember your conversation while it answers. And it must be GPU memory specifically — regular RAM is roughly ten times too slow to feed the chip.

MIT license · open weights

DeepSeek V4 Flash

284B parameters

An efficiency-first mixture-of-experts model: 284 billion parameters in total, but only 13 billion do the work on each word it produces. All 284 billion still have to sit in GPU memory — the model chooses which experts to wake per token, so every expert must already be loaded.

The memory math

284 billion parameters × 1 GB = 284 GB → + 20% working room = 340.8 GB of GPU memory

✓ Recommended

3× H200 NVL

423 GB GPU memory vs 340.8 GB needed

$96,000 total price

The minimum honest setup: 423 GB of NVLink-bridged memory against the 340.8 GB the model needs.

1,800 W — as much as about 1.5 average homes drawing power around the clock, and a day of running it uses about 48% of a 90 kWh electric-car battery.

Request this build — $96,000

DGX B300

2,304 GB GPU memory vs 340.8 GB needed

$550,000 total price

The room-to-grow pick: the same model with 6.7× headroom for more users, longer contexts, or a bigger model later.

14,000 W — as much power as about 12 average homes, and a day of running it drains about 3.7 full 90 kWh electric-car batteries.

Request this build — $550,000

MIT license · open weights

GLM 5.2

744B parameters

Z.ai's frontier open model: 744 billion parameters (about 40 billion active per token), MIT-licensed, with a one-million-token context window. One of the strongest open coding and agent models in the world — and it needs the better part of a terabyte of GPU memory to run.

The memory math

744 billion parameters × 1 GB = 744 GB → + 20% working room = 892.8 GB of GPU memory

✓ Recommended

DGX B300

2,304 GB GPU memory vs 892.8 GB needed

$550,000 total price

One box, no assembly, no risk: 2,304 GB of factory-fused NVLink memory against the 892.8 GB the model needs.

14,000 W — as much power as about 12 average homes, and a day of running it drains about 3.7 full 90 kWh electric-car batteries.

Request this build — $550,000

For comparison — not for sale as a build

10× RTX PRO 6000 (for comparison only)

960 GB on paper

$89,000 and it still wouldn't work well

Ten workstation cards do add up to 960 GB on paper — but these cards have no NVLink. They can only talk through the motherboard, hundreds of times slower than the model needs, so the pile holds the model and cannot usefully run it. We show the math because it is honest; we do not sell it as a build.

6,000 W — as much power as about 5 average homes, and a day of running it drains about 1.6 full 90 kWh electric-car batteries.

MIT license · open weights

GLM 5.1

754B parameters

The previous generation of Z.ai's flagship — slightly larger at 754 billion parameters, MIT-licensed, built for long agentic sessions with hundreds of tool calls. Companies pick it when their stack is already tuned for it.

The memory math

754 billion parameters × 1 GB = 754 GB → + 20% working room = 904.8 GB of GPU memory

✓ Recommended

DGX B300

2,304 GB GPU memory vs 904.8 GB needed

$550,000 total price

Same answer as GLM 5.2: at 904.8 GB needed, one DGX B300 is the smallest machine we sell that runs it properly.

14,000 W — as much power as about 12 average homes, and a day of running it drains about 3.7 full 90 kWh electric-car batteries.

Request this build — $550,000

For comparison — not for sale as a build

10× RTX PRO 6000 (for comparison only)

960 GB on paper

$89,000 and it still wouldn't work well

960 GB of paper capacity, no NVLink to fuse it. Honest math, not a real machine — see the cluster explanation below.

6,000 W — as much power as about 5 average homes, and a day of running it drains about 1.6 full 90 kWh electric-car batteries.

When a build combines machines

A cluster is several machines wired together tightly enough to work on one job as if they were one machine. Every combined build above shows its combined memory, power, and price. The wiring is the whole difference: H200 NVL cards share memory over NVLink bridges at 900 GB/s, and a DGX B300 fuses eight GPUs at 1.8 TB/s each — while desktop cards have no NVLink at all and can only chat through the motherboard. A pile of desktop cards adds up its memory on paper; a real cluster actually shares it.