Ming-Image-0.1-Design vs Qwen-Image 2.1: The Licence Is the Real Difference

Two open image models landed three days apart in September 2026 and, read off a spec sheet, they look like the same purchase. Ming-Image-0.1-Design, built by Ant Group's inclusionAI team, put its weights on Hugging Face on 17 September under an MIT licence. Qwen-Image 2.1, from Alibaba's Qwen team, followed on 20 September under the Qwen Research License Agreement. Both emit native RGBA, both claim 2K output, and both are aimed at the same buyer: a team that wants image generation it can host, inspect and modify itself. The two things that actually separate them are the two a spec sheet is worst at showing — what you are legally allowed to do with the result, and how much silicon it takes to produce one.

That is not a rhetorical framing. For one of these models the licence is a hard stop on commercial use; for the other it is a non-issue. And the parameter count each vendor prints on the box understates the download by a factor of three or four. Those two facts decide the purchase long before any benchmark does — and the benchmarks, as it happens, cannot be compared at all.

What each repository actually contains

Neither model is one network. Both are pipelines, and the pipeline is where the size lives:

• Generator — Ming-Image-0.1-Design ships a 6.15B-parameter design transformer (12.31 GB in BF16); Qwen-Image 2.1 ships a 7.1B single-stream DiT with 32 layers, block-causal attention and an undisclosed number of active parameters (about 14.2 GB).

• Text side — Ming-Image-0.1-Design pairs its transformer with a 17.01B multimodal LLM (34.02 GB) and a 3.09B connector (6.17 GB); Qwen-Image 2.1 uses a Qwen3-VL 8B text encoder (about 17.5 GB).

• Total parameters — 26.44B for Ming-Image-0.1-Design against roughly 15–16B for Qwen-Image 2.1. Note that neither vendor's headline number is the total: "6B" and "7B" both describe the generator only.

• Repository size — 52.88 GB for Ming-Image-0.1-Design (49.25 GiB of that is weights); roughly 33 GB for Qwen-Image 2.1.

• Transparency — both output native RGBA straight from the model rather than via a post-hoc matting step. Ming-Image-0.1-Design uses its own RGBA VAE; Qwen-Image 2.1 uses a 64-channel RGBA VAE with 16× spatial compression.

• Default generation — Ming-Image-0.1-Design is validated at 2048×2048, 12 steps, BF16, CFG 1.0; Qwen-Image 2.1 defaults to 2048×2048 over 40 steps with seven aspect-ratio presets.

• Licence — MIT for Ming-Image-0.1-Design; the Qwen Research License Agreement, non-commercial, for Qwen-Image 2.1.

The "6B" label is doing a lot of work

Ant Group's repository is titled around a 6-billion-parameter model, and the transformer really is 6.15B. But the weights you download are 49.25 GiB, the repository is 52.88 GB, and the configuration the vendor validated — 2048×2048 at BF16 with the full text stack resident — is specified for one 80 GiB CUDA GPU. That is not a rounding error on a 48 GB card; it is a card you cannot use. A 48 GB accelerator will not hold the pipeline, and neither will two of them without offload gymnastics the vendor does not document.

Qwen-Image 2.1 is the more forgiving download, with the caveat that Alibaba publishes no official VRAM figures at all. The model card's only guidance is to enable CPU offload on smaller cards. What has been measured since launch is more useful than what has been published: an int8 pipeline runs around 16 GiB in total and fits a 24 GB card fully resident, while a full BF16 run has been reported anywhere from about 17 GB with CPU offload to roughly 40 GiB peak at 1024×1024 depending on how the text encoder is placed. Reported single-image latency ranges from a few seconds on a B200 to under ten on a current workstation card. Treat every one of those numbers as community measurement rather than specification — they vary with offload strategy by more than the models differ from each other.

The practical summary: Qwen-Image 2.1 is a 3090-class model if you are willing to offload; Ming-Image-0.1-Design is a datacentre-class model with no smaller configuration.

The licence is the fork in the road

This is the part that will actually stop a project, and it is the part the two releases treat most differently.

Ming-Image-0.1-Design is MIT. Ship it in a paid product, fine-tune it, redistribute the weights, keep your changes closed — none of that requires a conversation with Ant Group.

Qwen-Image 2.1 is not. It carries the Qwen Research License Agreement dated 20 September 2026, which licenses the weights for research and evaluation only. Commercial use needs a separate agreement from Alibaba, and no price for that agreement is published. Derivatives and redistributions have to carry the licence, a "Built with Qwen" or "Improved using Qwen" notice and the Hangzhou Tongyi Laboratory copyright line, and "Qwen" cannot be the primary name of a derived model. The licence is governed by Chinese law with jurisdiction in Hangzhou.

There is one genuinely clarifying point in the vendor's own follow-up: Alibaba has stated that generated images are not part of the licensed "Materials", so you keep the rights to what the model produces. The restriction attaches to the weights and the model, not to your output. That is a meaningful carve-out and it is worth reading precisely, because the easy summary — "the images are yours, the model is not" — is the correct one.

It is also a change of direction. Earlier Qwen-Image releases, including the original Qwen-Image and the 2511 and 2512 builds, remain Apache 2.0 and commercially permissive. A team that picked Qwen-Image last year for its licence and upgrades to 2.1 this month inherits a research-only restriction it never agreed to. If permissive terms are the requirement, Qwen-Image 2.1 is not the newer version of what you had — it is a different deal.

What each one is built to do

The licence split mirrors a difference in intent, and it is the difference that should drive the choice when both are legal options for you.

Ming-Image-0.1-Design is a design tool. The text stack exists to expand a short brief into an 8K structured prompt, and the vendor ships "skills" around it — a Design Skill that produces a visual draft in around 15 seconds, and a PPT Skill aimed at slide-shaped output. Its companion repository, Ming-Image-0.1-Design-Layer, takes a flattened design and returns it as separate layers, which is the one capability here nothing else in the open field offers. inclusionAI reports the Layer model running about 4.3× faster than a 20B open layer model on the same job (183 seconds against 795) — vendor-reported, unreproduced, and the comparison target is not named precisely enough to check.

Qwen-Image 2.1 is a generator-and-editor. One checkpoint does both text-to-image and instruction-based editing, accepts up to 10 reference images for a single edit, and supports local edits that leave the rest of the frame alone. It shipped with day-zero support in ComfyUI, Diffusers, vLLM-Omni, SGLang-Diffusion and LightX2V — a serving story Ming-Image-0.1-Design does not yet have, since its own serving guide points at vLLM-Omni.

Read that as a fork rather than a ranking. If the job is "produce a layered, editable design asset from a brief", Ming-Image-0.1-Design is the only one of the two that does it. If the job is "generate and then iteratively edit against references", Qwen-Image 2.1 does it in one model and Ming-Image-0.1-Design does not do it at all.

The benchmarks do not compare, and that is the honest answer

Both vendors have numbers. Neither number can be placed next to the other, and any article that does so is inventing a result.

Ming-Image-0.1-Design's evidence is an Artificial Analysis board: first place on the Text-to-Image Leaderboard filtered to open weights for UI/UX Design, at an Elo of 1,082, ahead of Ideogram 4.0 Quality at 1,052, Ideogram 4.0 at 1,015, HunyuanImage 3.0 Instruct at 1,005 and FLUX.2 [dev] at 1,000. The vendor also reports win rates of 67.4% on layout, 67.0% on complex composition and 66.7% on text rendering, and first place across 12 metrics on Crello. The leaderboard is a third-party instrument, which makes the Elo the stronger of the two kinds of evidence here; the win rates and the Crello sweep are vendor-reported and should be read as claims.

Qwen-Image 2.1's evidence is Qwen-Image-Bench, a benchmark the Qwen team built, on which it scores 60.28 and leads the open-source field ahead of Nano Banana 2.0 at 59.82 and GPT Image 1.5 at 59.65. The source material flags the benchmark as Qwen-built, and the top three sit inside one point of each other — a gap well within the noise of a benchmark whose author is also a competitor in it.

So: a third-party Elo on a UI/UX design subset, against a vendor-built general score. Different instruments, different tasks, different judges. No one has published a head-to-head run of these two models against each other, and on the evidence available nobody can. The useful reading is narrow and worth stating plainly — Ming-Image-0.1-Design has third-party corroboration on design work specifically, Qwen-Image 2.1 has vendor-reported strength on general generation and editing, and the overlap between those two claims is smaller than the headline numbers suggest.

Getting either one onto an endpoint

Neither model is on OrcaRouter's catalogue today. Ming-Image-0.1-Design and Qwen-Image 2.1 are both self-host propositions right now: the weights are downloadable and the serving guides are real, but you are standing up the GPU yourself, and in Ming's case that GPU is an 80 GiB one. We list them here as open-weight releases, not as routes.

What is worth knowing before you commit hardware is what the same workload costs you as a call. Our routed image models include google/gemini-2.5-flash-image, google/gemini-3-pro-image-preview, google/gemini-3.1-flash-image-preview, the three Imagen 4.0 tiers (fast, standard and ultra), grok/grok-imagine-image and the GPT-Image family — openai/gpt-image-1, openai/gpt-image-1-mini, openai/gpt-image-1.5 and openai/gpt-image-2. Running a design brief against one of those for a week tells you whether Ming-Image-0.1-Design's layer output is a capability you actually need, at a fraction of the cost of the 80 GiB card, and it does so before you find out whether the licence or the VRAM was going to be the blocker. When you do move a self-hosted model into a production path, the same endpoint is where the routing lives — one key across 200+ models, automatic failover between providers, and 0% markup, so a vendor price cut is live on our side the same day it is announced.

Who should pick which

Pick Ming-Image-0.1-Design if the deliverable is a design asset and not a picture: layered output, structured prompt expansion, slide-shaped generation, and a licence that lets you sell whatever you make with it. Budget for a serious card and accept that the serving ecosystem around it is thinner than the Qwen side. It is also the more useful of the two if you need to put a number in front of a customer, because its strongest evidence is the only third-party number in this comparison.

Pick Qwen-Image 2.1 if the workflow is generate-then-edit, if you need reference-image conditioning, or if you want to run on hardware you already own. It is smaller, better served on day zero, and its licence is fine for the research and evaluation work most people actually do first. It becomes the wrong pick the moment the output is commercial and the legal review is real, because the only route to a commercial licence is a direct agreement with Alibaba and there is no published price to plan against.

Pick neither, yet, if what you need is image generation in production this month. Both of these are weights, and weights are a hardware and legal commitment before they are a capability. The design question — does layered output change what my team can ship — can be answered on a hosted endpoint first, and that answer is what should decide whether the 80 GiB card gets bought.

What to watch

Three things will move this comparison. Whether Ming-Image-0.1-Design gets a serving path beyond its own vLLM-Omni guide, because that is currently the gap between it and Qwen-Image 2.1's five-framework day-zero list. Whether Alibaba attaches a price to the commercial Qwen licence, because an unattached price is the single largest unknown in this pairing. And whether anyone — a lab, an independent evaluator, or a vendor with nothing to lose — runs these two against each other on the same prompts. Until that happens, the closest thing to a fair comparison is not a score at all. It is the licence line, and Ming-Image-0.1-Design wins that one outright.