Hy Image3.5 vs Ming-Image-0.1-Design: You Can Rent One and You Cannot Buy the Other

Hy Image3.5 and Ming-Image-0.1-Design went live five days apart in September 2026 and could not be more differently distributed. Tencent's model is a metered API: 2K images at $0.024 each on Tencent Cloud International, a preview that opened on 22 September, reachable today through Tencent's own products and cloud surfaces. inclusionAI's Ming-Image-0.1-Design is a repository: MIT-licensed weights on Hugging Face since 17 September, roughly 71.5 GB of BF16 safetensors, no endpoint, no price, no provider, and a Quick Start that points at a companion repository returning 404. One is a product you can call this afternoon. The other is a file you can download this afternoon.

That is the comparison, and it is not a quality contest — neither model has an independent score, so neither can be ranked on output. What can be compared is what each one lets you do, and where each one stops. Read that way, the two models are almost exact opposites on every axis a production team cares about.

The one dimension where the open model wins outright

Start with the licence, because it is the only place the comparison is not close.

Ming-Image-0.1-Design is MIT. Download the weights, fine-tune them on your own catalogue, ship the result inside a commercial product, host it for customers, keep the derived model — no per-call fee, no usage terms to negotiate, no vendor that can deprecate the endpoint under you. For a team with a privacy constraint, a fixed GPU budget, or a need to train on proprietary design assets, that is not a preference, it is the difference between having a route and not having one. No hosted image API, Tencent's included, can offer it.

What that licence costs is the rest of the article.

Everything else, one line at a time

• Access — Hy Image3.5 is live through Tencent Cloud TokenHub, Media Processing Service and Video on Demand, plus Yuanbao, ima, Miora, WorkRally, OnSolo and WorkBuddy, with a free window on three of those until 7 October 2026; Ming-Image-0.1-Design has no hosted endpoint anywhere, no inference provider deployment, and its own model card tells you to clone an inference repository and run a script.

• Cost — $0.024 per delivered 2K image, with failed and filtered generations unbilled and no input meter, against free weights and an 80 GiB CUDA GPU you have to find. The second number is not zero; it is the rental or depreciation on a card in that class, plus the engineering hours to stand the stack up.

• Output — Hy Image3.5 reaches 2K with the standard aspect ratios and no 4K tier; Ming-Image-0.1-Design recommends 2048×2048 or 1024×1024 for speed, at 12 sampling steps and CFG 1.0 in BF16.

• Transparency — Ming-Image-0.1-Design emits RGBA with transparent backgrounds as a native output, which is a genuine capability for icon, sprite and layered-composite work; Hy Image3.5 has no documented transparent-background mode.

• Conditioning — Hy Image3.5 takes up to five reference images per request and retains earlier turns across a multi-turn editing session; Ming-Image-0.1-Design's conditioning story is its text encoder, a multimodal LLM in the pipeline, with prompt enhancement expected from sibling models rather than built in.

• Deployment support — a managed API with uptime attached, against a vLLM-Omni path whose install guide resolves while its model-specific recipe does not.

• Independent evidence — nothing, on either side.

The evidence problem, which is worse than it looks

Both models launched without a third-party measurement, but they failed differently, and the difference is instructive.

Tencent's claim is a blind GSB preference study run with several hundred of its own professional designers, in which Hy Image3.5 beats Hy Image3.0 by about 30%, matches Seedream 5.0 Pro, and edges past Nano-Banana Pro and Qwen-Image-3.0 Pro. The methodology is sound, the participants work for Tencent, the prompt set is unpublished, and the result is a number nobody outside the company can re-run. On the Artificial Analysis text-to-image board read on 23 September 2026 there is no Hy Image3.5 row at all; the previous generation, HunyuanImage 3.0 Instruct, sits 65th of 156 at 964 Elo.

Ming-Image-0.1-Design's claim is a leaderboard graphic published inside its own repository, headed "Text to Image Leaderboard: UI/UX Design" with the footer "Elo scores from blind preference votes in our Image Arena" and an "Open Weights Leaderboard" badge. It places the model first at 1,082 Elo, above Ideogram 4.0 (Quality) at 1,052, Ideogram 4.0 at 1,015 and HunyuanImage 3.0 Instruct at 1,005. The graphic carries Artificial Analysis branding; the model does not appear on Artificial Analysis's live boards. We could not find the board it depicts on the public web, and it shows no vote count.

So Tencent's number at least describes a study the company ran. Ming's describes a study that may not exist outside its own repository — which is a harder problem, because it is not just unaudited, it is unlocatable. Neither should be quoted as a benchmark, and the honest position on both is that the quality question is open.

Why the architecture is worth a look anyway

One detail in Ming-Image-0.1-Design's pipeline is the most interesting technical thing in either announcement, and it is visible in the repository rather than in any press material. A third-party vLLM-Omni commit wires a ByT5 byte-level glyph encoder into a pipeline named `ming_flash_omni`, extending the tokenizer vocabulary with font and colour markers. That is a design intent you can read directly: the model appears to treat rendered text as glyphs with typographic attributes rather than as an emergent property of the prompt, which is the failure mode behind almost every badly-rendered poster an image model has ever produced.

It is also consistent with the model's stated purpose — UI, infographics, posters and other text-heavy design work — and with the recommended 2048×2048 output, since legible type needs pixels. But it is intent, not evidence. The repository layout shows six component folders with no `model_index.json`, meaning the model is a multi-stage pipeline rather than a single call, and the glyph-encoder wiring lives in someone else's inference framework rather than in the model card. Until a third party runs it and shows the text, it is an architectural claim.

Hy Image3.5's equivalent claim is more conventional and equally unverified: Tencent says in-image Chinese and English lettering improved, alongside better prompt-specified layout and stronger editing on real photographs. Also vendor-reported, also unreproduced.

How to actually use either one

The practical answer differs completely depending on which model you picked, and neither answer involves a leaderboard.

If you want Hy Image3.5, use the free window. Miora, WorkRally and OnSolo are open at no cost from 22 September to 7 October 2026, which is long enough to run your real prompts against it and form an opinion that Tencent's study cannot give you. Then decide whether a preview you cannot self-host, cannot fine-tune, and cannot measure independently belongs in a production path — and if it does, keep the generation call behind an interface you can repoint, because a preview carries no availability commitment and the price or the model can move under you.

If you want Ming-Image-0.1-Design, the first task is not generation, it is deployment. Budget for a single 80 GiB CUDA GPU, expect to clone an inference repository rather than call a pipeline, and check the vLLM-Omni recipe link before you plan around it, because it did not resolve when we looked. What you get on the other side is the only one of these two models you can fine-tune, the only one that emits a transparency channel, and the only one that cannot be taken away from you.

That last property is also the reason a routing layer and an open-weights model are complements rather than alternatives. OrcaRouter is the managed half — one key across more than 200 models, provider list price passed through at 0% markup so a vendor's rate change is live the same day, and automatic failover so an unproven preview can sit behind a production call without being a single point of failure. It is not the half that gives you weights. To be precise about our catalogue: neither Hy Image3.5 nor Ming-Image-0.1-Design is routable here. OrcaRouter hosts no Tencent image model and no inclusionAI model, and the hosted still-image models we do serve are Google's Nano Banana 2 (as google/gemini-3.1-flash-image-preview), GPT-Image-2, Imagen 4 Ultra and Grok Imagine Image.

What each one is for

Hy Image3.5 is for teams that want a picture cheaply and do not want to run anything: 2K output, per-image billing with no charge for discarded attempts, a working API today, and a quality claim you will have to test yourself. It is the lowest-friction option in the category and it comes with no ownership at all.

Ming-Image-0.1-Design is for teams whose constraint is not price but control — a licence they can build a business on, a model they can fine-tune, a transparency channel no hosted API is offering, and a GPU bill instead of a per-call one. It comes with no support, no endpoint, and a deployment story that is half finished.

The two do not compete for the same slot, and the reason is not quality. It is that one of them is a service and the other is an asset, and the question of which you need is answered before you ever look at a benchmark.