Ming-Image open weights or a hosted image API: what you operate

Open weights mean you run the model; a hosted API means you send a request. inclusionAI's Ming-Image-0.1-Design and Ming-Image-0.1-Design-Layer, released 2026-09-17, are two 6B-parameter MIT-licensed models; the Layer model outputs RGBA layers and the pair is reported to run on a single 80 GiB GPU (digitalapplied tracker, read 2026-10-02).

Sume is the hosted side. It serves the image models in its catalog through one API and does not host Ming weights or give you weights to run, per the docs I read on 2026-10-02.

What do the Ming models do?

Per the tracker, Design is a design-focused image model and Design-Layer decomposes a flat design into editable layers with transparent RGBA outputs. Both are MIT licensed. I read only the tracker, not the model cards, so check the licence and hardware notes on the model page before you plan around them.

What does a hosted call replace?

You skip the GPU, the model server, and the update work. What you take on instead is the request contract: Sume jobs are durable, a 202 job envelope arrives when a generation outlasts the 30-second sync wait, and image generation bills only when it completes.

Results come back as Sume-hosted signed URLs under data[].url, so you store the Sume URL rather than a raw provider URL.

Which one fits which job?

Choose by the output you need. If you need true editable layers today, a layer model is the match, and Sume's route is a workaround: separate calls per layer, as in editable layers with Pillow. If you need flat images, edits, references and masks behind one key and one billing line, the Image API is that. Pilot both on ten real designs before you commit.