MOSS-Image Turbo — MLX INT4 package for LocalMuse

This package redistributes converted model weights under the Apache License,
Version 2.0 (see LICENSE), except the UltraReal adapter, which is distributed
under its own Civitai model license (below). It is a format conversion and is
not affiliated with
or endorsed by the OpenMOSS Team, lodestones, akhaliq, vizsumit, Alibaba Cloud or
Black Forest Labs.

MOSS-Image-SFT
  Copyright OpenMOSS Team. Apache License 2.0.
  https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT
  revision fc5de0d56fc910cbca6f6625a70f5f582b26deb1

  Includes the Qwen3.5 text encoder (Copyright Alibaba Cloud / Qwen Team,
  Apache License 2.0) and the FLUX.2 small decoder (Copyright 2026 Black Forest
  Labs, Apache License 2.0) as distributed in that repository.

MOSS-Image-Base 8-step DMD2 adapter (lora_dmd2.safetensors)
  Copyright lodestones. Apache License 2.0.
  https://huggingface.co/lodestones/MOSS-Image-Base-8-steps-dmd
  revision da05cb3b52899179e2bb5f7efe1058fefcbae21b

MOSS-Image-SFT 4-step turbo adapter v2.0 (lora.safetensors)
  Copyright akhaliq. Apache License 2.0.
  https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps
  revision 2ded95c37a029360245926fe72e475e30039bbfb

UltraReal MOSS v2 (ultra_real_moss_v2.safetensors)
  By vizsumit, distributed under the Civitai model license of
  https://civitai.com/models/2462105?modelVersionId=3416438 (commercial use and
  derivatives permitted). SHA-256 b7574656d90d668bde19c809812cee5f342433533173e34416f6177cd736ba4e

Modifications: tensors were renamed and converted to affine INT4 (group 64),
BF16 and FP32 for MLX as described in README.md; the adapters are stored as BF16
factors (UltraReal's fused qkv split into q, k, v) and applied without merging.
No weights were retrained.