Question 1

Can the NVIDIA H100 80GB PCIe run qwen3 32B?

Accepted Answer

qwen3 32B needs about 21GB working set (weights 20GB + KV cache 1GB at 4k context). The NVIDIA H100 80GB PCIe offers ~80GB usable — verdict: full. 21.0 GB needed of 80 GB usable — headroom for context

Question 2

How fast is qwen3 32B on the NVIDIA H100 80GB PCIe?

Accepted Answer

~47–77 tok/s on the NVIDIA H100 80GB PCIe. computed roofline: 2039 GB/s × 0.46–0.76 efficiency window / 20.0GB (Q4_K_M) — bands, never points.

Question 3

What should I run qwen3 32B on instead?

Accepted Answer

Tesla V100 32GB (used) runs it fully ($640).

Can the NVIDIA H100 80GB PCIe run qwen3 32B?

Run qwen3 32B on instead