DeepSeek-V4-Pro
The biggest model in our store: 1.6 trillion parameters. Even though it only "actively thinks" with 49 billion of them per word, every one of the 1.6 trillion has to be held in memory somewhere.
📊 Arena.ai Text Leaderboard, rank ~38, Elo 1457 — filtered to Open Source models. · source
The memory math
parameters = 1600 billion
× 1 GB per billion = 1600 GB (raw weight size)
× 1.2 (20% working room for the model to actually run, not just sit in storage)
= 1,920.0 GB minimum memory needed
× 1 GB per billion = 1600 GB (raw weight size)
× 1.2 (20% working room for the model to actually run, not just sit in storage)
= 1,920.0 GB minimum memory needed
Why the 20% headroom? A model doesn't just sit in memory — while it's answering you, it needs extra scratch space to hold the conversation so far and its in-progress calculations. Buy exactly the raw size and it will crash under real use.
Two ways to get there
Option A
2× NVIDIA DGX B200$1,030,820Total memory2,880 GB
Total power28.6 kW
Total price$1,030,820
One DGX B200 (1,440 GB) isn't enough on its own — this model needs 1,920 GB. Two DGX B200 systems networked together (2,880 GB total) is the minimum real cluster that fits it.
Request this Quote✅ Recommended: Option B
1× NVIDIA GB300 NVL72$6,500,000Total memory20,000 GB
Total power135.0 kW
Total price$6,500,000
A single GB300 NVL72 rack (~20,000 GB) swallows this model with room for roughly 10 copies at once — the choice for a company planning to serve this model at real scale.
Request this QuoteWhat that power means
🏠
Running the recommended build 24/7 draws roughly 112.5 average homes' worth of continuous power.
🔋
One hour of running it uses about 150.0% of a typical 90 kWh electric-car battery.