Nami Tech Hardware

Model Advisor

Pick a big open-source model, and we'll show the memory math and the minimum real hardware you'd need to buy to run it.

8B 70B 405B
The math we use for every model

A model's size is its parameter count — the "B" in a name like Llama 3.1 405B means 405 billion parameters. The rule of thumb the whole industry uses: each billion parameters needs about 1GB of GPU memory, plus about 20% extra working room for the running conversation itself (this is called the "context" — the text being read and generated). So: minimum GB = parameters (billions) × 1GB × 1.2

Llama 3.1 405B

Meta · Apache 2.0 — open source, free to self-host
Parameter count
405B
405 billion parameters × 1GB = 405GB just to hold the model. Add 20% working room for context and you get the minimum any real setup must clear:
405 × 1GB × 1.2 = 486GB minimum GPU memory
Below that number, the model cannot even load — it's not a "runs slow" problem, it's a "does not run" problem.
Minimum real setup that clears 486GB
1 × DGX H100 (640GB each) $200,000
Combined memory
640GB
Clears the 486GB minimum with 154GB to spare.
Combined power
10,200W
about 8.5 average homes running non-stop.
Total price
$200,000
Everything needed, nothing extra.

One DGX H100 supplies 640GB — well past the 486GB minimum, leaving ~154GB of headroom for longer conversations or future growth.

Why this hardware, not something elseWhy one DGX H100 and not, say, 6 separate H100 cards? A single DGX H100 is already 8 H100s pre-wired with NVLink into one 640GB machine — it clears the 486GB minimum in one purchase, one support contract, and one power cable, instead of you having to buy 6+ individual cards and build the fast interconnect between them yourself. For a model this size, buying the packaged system is simpler and just as capable.
Request this build for Llama 3.1 405B

Qwen3 235B

Alibaba · Apache 2.0 — open source, free to self-host
Parameter count
235B
235 billion parameters × 1GB = 235GB just to hold the model. Add 20% working room for context and you get the minimum any real setup must clear:
235 × 1GB × 1.2 = 282GB minimum GPU memory
Below that number, the model cannot even load — it's not a "runs slow" problem, it's a "does not run" problem.
Minimum real setup that clears 282GB
4 × A100 80GB (80GB each) $48,000
Combined memory
320GB
Clears the 282GB minimum with 38GB to spare.
Combined power
1,600W
about 1.3 average homes running non-stop.
Total price
$48,000
Everything needed, nothing extra.

4 A100 80GB cards, wired together, supply 320GB — past the 282GB minimum, at a lower price and lower power draw than H100-based alternatives.

Why this hardware, not something elseWhy A100 and not the newer H100? Because H100's extra speed mostly helps with training new models and serving many users at once — it doesn't add any GPU memory. Since the only hard requirement here is clearing 282GB, 4 A100s do it for $48,000 and 1,600W, versus roughly double the price and power for the same memory in H100s. We only recommend the faster (pricier) card when the workload actually needs the speed.
Request this build for Qwen3 235B

DeepSeek-V3 671B

DeepSeek · MIT — open source, free to self-host
Parameter count
671B
671 billion parameters × 1GB = 671GB just to hold the model. Add 20% working room for context and you get the minimum any real setup must clear:
671 × 1GB × 1.2 = 805.2GB minimum GPU memory
Below that number, the model cannot even load — it's not a "runs slow" problem, it's a "does not run" problem.
Minimum real setup that clears 805.2GB
2 × DGX H100 (640GB each) $400,000
Combined memory
1,280GB
Clears the 805.2GB minimum with 474.8GB to spare.
Combined power
20,400W
about 17 average homes running non-stop.
Total price
$400,000
Everything needed, nothing extra.

One DGX H100 (640GB) is not enough on its own. Two DGX H100 systems wired together supply 1,280GB — comfortably past the 805GB minimum.

Why this hardware, not something elseWhy 2 DGX H100 systems? One alone (640GB) is short of the 805GB this model needs — it would fail to load, full stop. Two is the smallest number of complete, real-cluster systems that clears the bar (1,280GB). We didn't recommend the GB200 NVL72 rack here: it would also work (13,500GB), but at $6.5M for a single rack it's roughly 16x the price for far more capacity than this one model needs — the honest minimum is 2 DGX H100 systems.
Request this build for DeepSeek-V3 671B