Ming-Image-0.1-Design is an open-source series from inclusionAI for visual-design generation and editable layer decomposition, released under the MIT license on GitHub with companion weights on Hugging Face and ModelScope. The series includes two 6B-parameter models: Ming-Image-0.1-Design generates complete visual designs for UI, infographics, posters, and text-rich compositions, including RGBA output with transparent backgrounds; Ming-Image-0.1-Design-Layer decomposes a flattened design image into a requested number of independently editable transparent RGBA layers, using an input image plus a per-layer specification. Both checkpoints run through a single infer.py CLI that accepts a local checkpoint directory or a Hugging Face Hub repo ID, resolved to one immutable local snapshot before loading. Sampling defaults are task-specific: text-to-image runs 12 steps at CFG 1.0 across 1024 or 2048 resolution buckets (2048 default), while layer decomposition runs 12 steps at CFG 2.0 across 512 or 1024 buckets (1024 default, preserving the input image's aspect ratio); a validate-only mode checks the checkpoint contract without loading weights. The minimum validated deployment is one GPU with at least 80 GiB of memory in BF16, with the CLI defaulting to eager attention and optionally accepting FlashAttention 2. For text-to-image, an optional prompt-rewriting step lets a VLM such as Ling-3.0-flash-VL or qwen3.8-27B expand a short caption into a Figma-style JSON layout with normalized coordinates and color specs, preserving each user-supplied text string verbatim; transparent-background generation requires starting the prompt with one of ten fixed RGBA trigger phrases. Layer decomposition uses an explicit per-layer specification rather than a free-form caption; the layer count is parsed from a "Decompose this image into N layers" line, or generated automatically via a num-layers flag. The layer model always returns the requested layers plus one leading composite image, which the CLI drops before saving the standalone RGBA layers as layer_01.png, layer_02.png, and so on. The project recommends serving both models through the vLLM-Omni inference framework.