Open-source model guide

Which model, and what it takes to run it

Three real open-source models from the arena.ai agent leaderboard — and the honest, no-hand-waving memory math for each one.

What "parameters" means

A model's size is its parameter count — the number of learned weights inside it. The "B" in "1600B" means billions. More parameters generally means a more capable model, but every one of them has to physically fit in GPU memory to run.

Why 1 GB per billion

These models are served at 8-bit precision, which is 1 byte per parameter. So a billion parameters is about a billion bytes — roughly 1 GB. That's the rule of thumb this whole store is built on: parameters (in billions) × 1 GB.

MoE and "active" params

These are Mixture-of-Experts models: only a fraction of the parameters ("active" params) do the work on any given token, which makes them fast. But the whole model still has to be loaded into memory — so memory is sized by total params, not active ones.

The +20% working room

Weights aren't the only thing in memory. The running model also needs room for the conversation it's processing (the KV cache) and scratch space. We add 20% on top of the raw weights so the recommended hardware actually works in practice, not just on paper.

At a glance

All three models side by side, with the minimum memory each one needs to run.

ModelTotal paramsActive paramsLicenseMin. memory (weights +20%)Recommended build
DeepSeek V4 Pro 1,600B 49B MIT 1,920 GB 3× DGX H100
GLM 5.2 Max 753B 40B MIT 904 GB 2× DGX H100
Nemotron 3 Ultra 550B 55B OpenMDW-1.1 660 GB 2× DGX H100

The math, model by model

DeepSeek V4 Pro

MIT Mixture of Experts

A 1.6 trillion parameter mixture-of-experts model — only 49B parameters are active per token, but every parameter still has to live in GPU memory.

1,600B Total parameters
49B Active per token
1,920.0 GB Minimum memory needed
1,920 GB Recommended build provides
Weights: 1,600B params × 1 byte (1 GB per billion) = 1,600 GB
+ 20% working room (KV cache + scratch) = 320 GB
Minimum GPU memory to run it = 1,920.0 GB
Needs 1,920 GB Build provides 1,920 GB (0 GB headroom)
100% used

Minimum setup: Startup Build — DeepSeek V4 Pro

3 DGX H100 systems, networked together, give 1,920GB of pooled GPU memory — exactly enough headroom for DeepSeek V4 Pro's 1,920GB minimum.

$1,125,000 Total build price

GLM 5.2 Max

MIT Mixture of Experts

A 753B parameter MoE model with 40B active parameters — a strong open model at roughly half the memory footprint of DeepSeek V4 Pro.

753B Total parameters
40B Active per token
903.6 GB Minimum memory needed
1,280 GB Recommended build provides
Weights: 753B params × 1 byte (1 GB per billion) = 753 GB
+ 20% working room (KV cache + scratch) = 151 GB
Minimum GPU memory to run it = 903.6 GB
Needs 904 GB Build provides 1,280 GB (376 GB headroom)
71% used

Minimum setup: Startup Build — GLM 5.2 Max

2 DGX H100 systems give 1,280GB of pooled memory, comfortably above GLM 5.2 Max's 903.6GB minimum, at the lowest real unit count that clears the bar.

$750,000 Total build price

Nemotron 3 Ultra

OpenMDW-1.1 Mixture of Experts

NVIDIA's 550B parameter MoE model, the smallest of the three but still far beyond what a single GPU — or even a single desktop card cluster — can hold.

550B Total parameters
55B Active per token
660.0 GB Minimum memory needed
1,280 GB Recommended build provides
Weights: 550B params × 1 byte (1 GB per billion) = 550 GB
+ 20% working room (KV cache + scratch) = 110 GB
Minimum GPU memory to run it = 660.0 GB
Needs 660 GB Build provides 1,280 GB (620 GB headroom)
52% used

Minimum setup: Startup Build — Nemotron 3 Ultra

Nemotron 3 Ultra needs 660GB minimum — just over what a single DGX H100 (640GB) provides, so the cheapest safe build is 2 DGX H100 systems (1,280GB).

$750,000 Total build price

Ready to match hardware to your model?

Our guided path takes about 30 seconds and ends with a priced build.

Help Me Choose