Home AI BenchMeasured on real home PCs

Z-Image Turbo did not fit a 16 GB card in our setup

Failed, then fixed2026-10-23 · Z-Image Turbo (bf16) · RTX 5060 Ti 16GB, RTX 3090 24GB

Portrait of an elderly man generated with Z-Image Turbo
Z-Image Turbo, 8 steps, made on the RTX 3090 24 GB node after the 16 GB card timed out.

Z-Image Turbo (Apache-2.0) is a 12.3 GB model plus an 8 GB text encoder.

What happened

  • On an RTX 5060 Ti 16GB node, requests ran past the gateway's time limit (about 90 seconds) and timed out. Our guess is that the model and encoder do not both fit in 16 GB at bf16, so the node kept moving weights in and out of VRAM (cause not confirmed).
  • Moving to an RTX 3090 24GB node with 32 GB of system RAM fixed it: 9.4 s per 1024×1024 image, 8 steps.

Takeaway

  • For Z-Image Turbo at bf16, plan for a 24 GB card. A quantized version may fit 16 GB (not tested).