Ming-Image-0.1-Design — MLX (8-bit DiT, 4-bit text encoder)
An MLX conversion of inclusionAI/Ming-Image-0.1-Design for mflux — 17.3 GB in total. Ming-Image is a design-focused text-to-image model (posters, cards, UI, typography) that outputs RGBA images.
| Component | Precision | Size |
|---|---|---|
Ling-mini-2.0 MoE text encoder (mllm/) | 4-bit (routers, norms, query tokens full precision) | 9.0 GB |
Qwen2 connector (connector/) | 8-bit | 1.4 GB |
Projection heads (mlp/) | bf16 | 0.06 GB |
DiT (transformer/) | 8-bit | 6.5 GB |
RGBA VAE (vae/) | bf16 | 0.25 GB |
The checkpoint's Qwen2.5 ViT, lm_head and audio router are not used for text-to-image and
are not included.
Usage
Requires mflux with Ming-Image support (the ming-image model family; until it is merged upstream,
install mflux from the branch that adds it).
mflux-generate-ming \
--model path/to/this/repo \
--prompt "A modern tech conference poster titled 'MLX SUMMIT 2026' ..." \
--width 1024 --height 1024 --seed 42
Defaults follow the official pipeline: 12 steps, guidance 1.0, flow-match Euler with the
checkpoint's static shift of 6. PNGs keep the alpha channel (--flatten-alpha for RGB).
Performance
The DiT stage is identical to the 5-bit-text-encoder variant: on a base M4 Mac mini (24 GB) 512² takes ~1 min (9.8 GB peak) and 1024² 4–4.5 min (14.5 GB peak). The text encoder is freed after encoding the prompt.
Fidelity
The MLX implementation was checked stage by stage against tensors dumped from the official PyTorch pipeline: DiT, VAE and connector at cosine similarity ≥ 0.9999, and the text encoder within the official model's own eager-vs-flash-attention spread (this includes reproducing the official pipeline's autocast-bf16 MoE routing exactly). Smallest variant. Typography stays clean, but compared with the official pipeline from the same noise, scene details (background, framing) can drift; prefer the 5-bit text encoder when it fits.
The DiT is deliberately kept at 8 bits: at 4 bits large headline lettering degrades.
License
MIT, as the original model. Credit to inclusionAI for Ming-Image.