Z-Image-Turbo by Alibaba Tongyi Lab, licensed under the Apache License 2.0.

The diffusion transformer is the INT8-ConvRot quantization published by Martin
Rizzo, also under the Apache License 2.0. Every layer is copied byte for byte.

Modifications by H3ddle: the weights were repackaged from the released
safetensors into one file per subsystem, the int8 matrices were transposed to
input-major so the GPU reads them coalesced, and the autoencoder's encoder half
was dropped because generating an image never runs it. The final layer's two small
linears were dequantized back to bf16 so that the head can run without a second
copy of the transformer; this reverses quantization rather than applying it, and
changes 1.2M of 6.15B parameters. Nothing was retrained, merged, or pruned, and
no weights were quantized here.