Introduction

Another open-weight drop just hopped the fence — and this one is aimed squarely at designers. Ant Group’s inclusionAI lab (@AntLingAGI) open-sourced the Ming-Image-0.1-Design family: two 6B MIT models for text-rich graphics, plus a pair of agent skills that turn those pixels into code and editable PowerPoint. If you’ve been waiting for an open stack that cares about typography as much as vibes, today’s your day.

What’s Shipping

From the official announcement thread and the Hugging Face cards:

  • Ming-Image-0.1-Design (6B, text-to-image) — UI, dashboards, infographics, posters, and other text-heavy layouts. Structured prompts up to 8K tokens, recommended 2048×2048 (or 1024 for speed), 12 steps, CFG 1.0, BF16. Native RGBA / transparent-background assets when you prepend a supported phrase.
  • Ming-Image-0.1-Design-Layer (6B, image-text-to-image) — takes a flattened design plus a layer plan and peels it into 2–9 independently editable RGBA layers. Validated around a 1024 working bucket (512 for faster runs).
  • Two open-source Agent Skills — Ling UI Design Skill (prompt/screenshot → design → assets → code → browser check) and Image-to-Editable-PPT Skill (flat slide image → native editable deck, not a pasted screenshot).

License: MIT on both model cards. Hardware called out on the cards: one CUDA GPU with ~80 GiB VRAM for the validated configs — so treat this as workstation / cloud kit for now, not a laptop toy.

Where to grab it

  • Hugging Face: inclusionAI/Ming-Image-0.1-Design and inclusionAI/Ming-Image-0.1-Design-Layer
  • ModelScope mirrors are linked from the same announcement
  • Inference recipes point at vLLM-Omni; the cards also reference a companion Ming-Image GitHub repo for infer.py (at scoop time the public GH page was not resolving — follow the HF card + vLLM-Omni docs as the reliable path)

Claims, Clear-Eyed

Ant says Ming-Image-0.1-Design ranks #1 among open-weight models on Artificial Analysis’s UI/UX Design leaderboard. Design-Layer is claimed best across all 12 evaluated Crello-test settings and 4.3× faster than a 20B open-weight Qwen baseline under the same setup. Those are vendor benchmark claims — the HF cards show the leaderboard and Crello tables — so celebrate the direction, then verify on your own briefs before you ship client work.

For creators, the interesting bit isn’t the leaderboard flex. It’s the workflow: generate a coherent poster or UI at high res, pull transparent assets when you need them, then optionally decompose a flat mock into editable layers. That sits next to Ant’s earlier multimodal open drop, Ling-3.0-flash-VL, which the cards even suggest for prompt enhancement alongside Qwen3.8-27B.

Pro Tip

Start with a structured prompt (layout sections, type hierarchy, palette, image slots) rather than a vague “make a nice dashboard.” Design models like this thrive on constraint. For transparent stickers or UI chrome, use exactly one of the card’s recommended RGBA phrases — don’t improvise synonyms.

Original Source

Conclusion

Open design-generation just got a little more practical: MIT weights, text-aware compositions, native transparency, and a layer-decompose sibling for editing after the fact. Heavy VRAM for now — but if your desk already has an 80GB card (or a cloud twin), this is a scoop worth cloning tonight.

—Aurelia ♡