CoolFace
Modelpublic

YvesStudioSmelt/z-image-turbo-Q8-affine

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face
0likes5downloads
Model Card

Z-Image-Turbo — Q8-affine (MLX) — Whispall build

8-bit affine MLX quantization (group size 32) of Tongyi-MAI/Z-Image-Turbo (quantized from revision f332072aa78be7aecdf3ee76d5c247082da564a6, Apache-2.0).

This is the higher-fidelity fallback tier for image generation in Whispall, the local, privacy-first macOS dictation & assistant app. The default ship tier is the 4-bit sibling (see below); Q8 trades a larger download and more RAM for slightly crisper fine detail.

Contents

A complete Z-Image model directory:

transformer/     DiT weights (quantized) + config.json
text_encoder/    Qwen3-4B-based text-encoder weights (quantized) + config.json
vae/             VAE weights + config.json
tokenizer/       tokenizer files
scheduler/       scheduler_config.json
quantization.json   MLX quantization manifest (bits=8, group_size=32, mode=affine)

Format / compatibility

  • —MLX affine quantization — loadable by the WhispallIMG engine (Whispall's vendored fork of `mzbac/zimage.swift`, MIT) and by MLX-based loaders that understand the quantization.json manifest convention.
  • —Not loadable by diffusers / PyTorch — use the upstream bf16 repo for that.
  • —Apple Silicon only (MLX/Metal).

Reproduce

WhispallIMG quantize -i <Z-Image-Turbo bf16 dir> -o z-image-turbo-Q8-affine --bits 8 --group-size 32

License & attribution

  • —Weights: Apache-2.0, © the Z-Image-Turbo authors (Tongyi-MAI / Alibaba). This repo is a quantized derivative; the original license and copyright carry over unchanged.
  • —Quantization performed with MLX; engine code MIT (fork of mzbac/zimage.swift).

Sibling repo

  • —z-image-turbo-Q4-affine — 4-bit variant (Whispall's ship default: smaller download, ~6 GB steady RAM with the engine's staged/capped flags).