transformer: 20.43B DMD2 4-step DiT bf16 (byte-identical to upstream)

text_encoder: Qwen2.5-VL-7B (byte-identical to upstream)

vae: AutoencoderKLQwenImage (byte-identical to upstream)

model_index.json

scheduler: packaged static shift-3 FlowMatchEuler config

tokenizer: fast tokenizer.json converted from the upstream slow files (token-id verified)

LICENSE (Apache-2.0, additional information per upstream)

NOTICE: NVIDIA Open Model License attribution (OML 3.1)

README: MLX snapshot notes, distillation defaults, license notice

initial commit