Downloads · 30 days
0
roman220220/z-image-turbo-gptq-mlx-8bit
z-image-turbo-gptq-mlx-8bit is a text-to-image model from roman220220. Use it when you need an image from a text prompt. It is set up for mlx. The card lists the license as apache-2.0.
<p align="center" <img src="llmtray-banner.png" alt="LLMTray" width="100%" </p
Downloads · 30 days
0
Access
Public
Updated Sep 27, 2026
Repo size
11 GB
Likes
0
Public
Click a slice to open those files.
.safetensors11 GB · 100%
From the Hugging Face model README
▶ Run it locally in LLMTray
A free, native macOS app for local AI on Apple Silicon — chat, images, music, agents and an OpenAI-compatible API. Point it at this model; nothing leaves your Mac.
An 8-bit MLX quantization of Z-Image-Turbo
(Tongyi-MAI, 6.15B diffusion transformer, 9-step distilled), for
mflux on Apple Silicon — but using
GPTQ Hessian-based error correction instead of mflux's own
round-to-nearest (RTN) --quantize 8, so it's the same on-disk size and
speed as a plain mflux-save --quantize 8, with better-preserved weights.
--quantize 8Nothing about the format — same shard layout, same key names, same
mx.quantize(mode="affine") grid mflux itself uses. What's different is
how the quantized values were chosen: instead of rounding every weight to
the nearest grid point independently (RTN), each of the 30 transformer
blocks' attention (to_q/to_k/to_v/to_out.0) and feed-forward
(w1/w2/w3) linears was corrected with GPTQ — quantize one column, measure
the rounding error, spread it onto the not-yet-quantized columns weighted
by that layer's activation Hessian, repeat. The Hessian was estimated from
real activations captured while running the model's own denoising loop
(hooks on every target Linear, several prompts, all 9 steps) — not
token/image data, since this is a diffusion transformer, not an LLM.
Full method + code: zimage-quant
(sibling of nemotron-extreme-quant, which does the same GPTQ correction
for LLMs — gptq_nbit is shared, unmodified code between the two).
Same prompt, same seed (42), same 9 steps, RTN-8bit vs this GPTQ-8bit:
RTN (plain --quantize 8) | GPTQ (this repo) |
|---|---|
![]() | ![]() |
At 8 bits the two are visually near-identical (PSNR ≈ 35.3dB between them, mean abs pixel diff ≈ 2.2/255) — 8 bits alone already has enough precision that RTN's per-weight rounding error is small. The 8-bit release is mostly a correctness baseline and a drop-in you can trust; the more interesting quality delta shows up at 4-bit — see z-image-turbo-gptq-mlx-4bit and the smaller-but-more-stable z-image-turbo-gptq-mlx-mixed (attention 8-bit + feed-forward 4-bit, 6.3GB).
pip install mflux
mflux-generate-z-image-turbo \
--model /path/to/local/checkout/of/this/repo \
--base-model z-image-turbo \
--prompt "your prompt" --steps 9
Or use it directly from LLMTray's
built-in image generation (the model chooses to call generate_image
mid-conversation; LLMTray drives mflux under the hood) — point LLMTray's
image-model setting at this repo.
~10GB total (transformer + text encoder + VAE + tokenizer), same as plain
mflux-save --quantize 8 — GPTQ's win here is quality-at-this-size, not a
smaller file.
Local-first AI tools for macOS by IPSupport — nothing leaves your Mac.
LLMTray uses this model for in-chat image generation (it drives mflux under the hood).
Licensed under the Apache License 2.0, the same license as the base model — see LICENSE.
Modified from Tongyi-MAI/Z-Image-Turbo: the attention (to_q/to_k/to_v/to_out.0) and feed-forward (w1/w2/w3) linears of all 30 transformer blocks were GPTQ-corrected and quantized to 8-bit, and the model was saved in mflux's MLX checkpoint format (same layout as mflux-save --quantize 8). The weights and configuration files in this repo are therefore modified versions of the original, not the original files.