Model advisory · sourced from arena.ai/leaderboard/agent, Open Source filter

How much memory does a model actually need?

A model's weights (its "brain") have to sit entirely in GPU memory while it runs — a model can't be partially loaded. The math is simple: parameters (in billions) × 1 GB, plus 20% working room for the temporary scratch space the model uses mid-conversation. Below, watch what happens when that number meets real hardware.

MIT

GLM-5.2

Z.ai

The math 744B × 1.2 = 893 GB

❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)

This needs 893 GB. This hardware is nowhere near enough — the bar below blows straight through the edge of the card.
capacity: 141 GB
6.3x over capacity — will not load

✅ Minimum that actually fits: 1x NVIDIA HGX H200 8-GPU Server (1,128 GB)

1,128 GB available for a 893 GB model.
usedcapacity: 1,128 GB
fits, with 235 GB of headroom to spare

Why this combination: One HGX H200 server already holds 1,128 GB — more than the 893 GB GLM-5.2 needs — so a single server is the smallest thing that actually fits it.

Total price $370,000
Total power 8,000 W
In homes about 7 homes running around the clock
See this as a startup recommendation →
MIT

DeepSeek V4 (Pro)

DeepSeek

The math 1600B × 1.2 = 1,920 GB

❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)

This needs 1,920 GB. This hardware is nowhere near enough — the bar below blows straight through the edge of the card.
capacity: 141 GB
13.6x over capacity — will not load

✅ Minimum that actually fits: 2x NVIDIA HGX H200 8-GPU Server (2,256 GB)

2,256 GB available for a 1,920 GB model.
usedcapacity: 2,256 GB
fits, with 336 GB of headroom to spare

Why this combination: A single HGX H200 server (1,128 GB) isn't enough for DeepSeek V4 Pro's 1,920 GB requirement. Two servers clustered together give 2,256 GB — enough room, at the lowest price that qualifies. A GB300 rack (20 TB) also fits, with far more headroom, at a much higher price.

Total price $740,000
Total power 16,000 W
In homes about 13 homes running around the clock
See this as a startup recommendation →
Modified MIT

Kimi K2.7 (Code)

Moonshot

The math 1000B × 1.2 = 1,200 GB

❌ If you tried: one NVIDIA H200 SXM 141GB (141 GB)

This needs 1,200 GB. This hardware is nowhere near enough — the bar below blows straight through the edge of the card.
capacity: 141 GB
8.5x over capacity — will not load

✅ Minimum that actually fits: 2x NVIDIA HGX H200 8-GPU Server (2,256 GB)

2,256 GB available for a 1,200 GB model.
usedcapacity: 2,256 GB
fits, with 1,056 GB of headroom to spare

Why this combination: One HGX H200 server (1,128 GB) is just short of Kimi K2.7's 1,200 GB requirement. Two servers clustered together give 2,256 GB, comfortably enough, at the lowest price that qualifies. A GB300 rack also fits, with far more headroom, at a much higher price.

Total price $740,000
Total power 16,000 W
In homes about 13 homes running around the clock
See this as a startup recommendation →