Hardware for Local AI

GPUs, prebuilt PCs, Macs and workstations — what they cost, how much VRAM they have, and which open models you can actually run on them.

The quiet path to 70B+ models: M5 Ultra 96 GB from $5,499 runs 70B at Q4 — slowly, but at all.

The fastest 24 GB card for local AI: 32B models at Q4 and 7–8B at 55–95 tok/s — but 70B still does not fit.

16 GB of VRAM at a mainstream price: 8–14B models at Q4 with room, at ~$575–900 as of October 2026.