Which model, and what does it need?
A model's size is its parameter count — the "B" in a name like
qwen3.5-397b means 397 billion. Every parameter has to sit in GPU memory, stored in
about 1 byte each (roughly 1 GB per billion parameters), plus about 20% extra "working room" for
the temporary math the GPU does while it writes each word. That's the whole rule:
Minimum GPU memory ≈ parameters (billions) × 1 GB × 1.2
All three below are genuinely open source (Apache 2.0 or MIT — check the license column yourself) and all are 100B+ parameters, picked from the world agent leaderboard.
Apache 2.0 · Alibaba
Qwen3-235B-A22B
Parameters235B
Minimum GPU memory282.0 GB
Math: 235B × 1 GB × 1.2 = 282.0 GB minimum.
Mixture-of-experts: only 22B parameters are "active" per word generated, but all 235B still have to sit in memory — active params save compute, not memory.
MIT · Zhipu AI
GLM-5
Parameters745B
Minimum GPU memory894.0 GB
Math: 745B × 1 GB × 1.2 = 894.0 GB minimum.
Mixture-of-experts: 44B active parameters per token out of 745B total.
MIT · DeepSeek AI
DeepSeek-V4-Pro
Parameters1,600B
Minimum GPU memory1,920.0 GB
Math: 1,600B × 1 GB × 1.2 = 1,920.0 GB minimum.
Mixture-of-experts: 49B active parameters per token out of 1,600B (1.6 trillion) total.
See the actual build
The minimum memory number above only tells you the floor. To see real hardware that clears it — with price and power attached — use the "help me choose" paths for a small startup or a mid-size company.