HP Z6 G4 tests local Qwen limits

This is the practical edge of local AI right now: used enterprise towers look powerful on paper, but VRAM still decides what models feel usable.

• –HP positioned the Z6 G4 as a real workstation platform with dual Xeon support, ECC memory, and room for professional GPUs, which makes it credible as a homelab inference box.
• –The Quadro RTX 6000's 24GB VRAM is the limiting factor here; it is better suited to smaller or quantized coding models than comfortable 70B-class local inference.
• –128GB of system RAM helps with CPU offload and experimentation, but once weights spill out of VRAM, speed and responsiveness usually fall off hard.
• –The clustering question is telling: budget buyers increasingly think in terms of chaining older boxes together, even though larger single-node GPU memory is usually the cleaner path for local LLM work.