Qwen3.5-397B-A17B
A mixture-of-experts model: 397 billion parameters live in memory, but only 17 billion "activate" per response — which is why it runs faster than its size suggests.
📊 Arena.ai Text Leaderboard, rank ~60, Elo 1443 — filtered to Open Source models. · source
The memory math
parameters = 397 billion
× 1 GB per billion = 397 GB (raw weight size)
× 1.2 (20% working room for the model to actually run, not just sit in storage)
= 476.4 GB minimum memory needed
× 1 GB per billion = 397 GB (raw weight size)
× 1.2 (20% working room for the model to actually run, not just sit in storage)
= 476.4 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
✅ Recommended: Option A
5× NVIDIA RTX PRO 6000 Blackwell Server Edition$65,000Total memory480 GB
Total power3.0 kW
Total price$65,000
5 cards × 96 GB = 480 GB, just over the 476.4 GB needed. This is a "pile of cards" cluster — cheapest way to fit the model, but the cards talk over standard networking, not NVLink.
Request this QuoteOption B
1× NVIDIA DGX B200$515,410Total memory1,440 GB
Total power14.3 kW
Total price$515,410
One DGX B200 has far more memory (1,440 GB) than this model needs, but it's a real NVLink-integrated system — faster and simpler to run, for about 8x the price.
Request this QuoteWhat that power means
🏠
Running the recommended build 24/7 draws roughly 2.5 average homes' worth of continuous power.
🔋
One hour of running it uses about 3.3% of a typical 90 kWh electric-car battery.