Hy Image3.5 vs Hy Image3.5 Preview: One Model, Two Names, and a Deadline

Search for Hy Image3.5 and you will find two things that look like two products. Search for Hy Image3.5 preview and you will find the same thing again, spelled slightly differently, sometimes with a hyphen, sometimes with a space, sometimes with a Chinese name attached. There is no separate stable release hiding behind the longer name. Every source that describes a current Tencent image model in this generation is describing one artefact — the model Tencent names in Chinese as the Hunyuan image 3.5 preview, served under the identifier hy-image-v3.5-preview — and the shorter name is what people call it when they are not being careful. This is a naming and availability question dressed up as a comparison, and the practical stakes are real: if you go looking for a stable version to pin a production pipeline to, you will not find one, because it has not been released yet.

What each name actually refers to

The long name is the accurate one. Tencent's own materials call the model a preview and serve it under a preview identifier, and the company's stated cadence for this line is preview first and official version later. The short name is what the model gets called in coverage, in forum posts, in the tool integrations and in the ComfyUI node list, because nobody types the word "preview" when they are naming a thing they are using.

This is not a cosmetic distinction, because the two names imply different answers to the question a team actually asks:

• Can I pin a production pipeline to it? — Not to a stable identifier. The only identifier in circulation is the preview one, and Tencent has published no date for a non-preview release.

• Is the short name a different, newer model? — No. There is no second artefact. A source that says "Hy Image3.5 launched" and a source that says "Hy Image3.5 preview launched" are reporting the same event on the same date.

• Which name should I search for? — Both. The short name finds commentary and integrations; the long name and the identifier hy-image-v3.5-preview find the documentation that actually tells you how to call it.

• Does the naming affect pricing? — No. Both names resolve to one rate card: ¥0.15 per 2K image domestically on Tencent Cloud and $0.024 an image on the international endpoint, billed on delivered output images only.

How you actually get it, in the order you will need it

The API surface is Tencent's own. The model is served through Tencent Cloud, reachable through the company's TokenHub API surface, and the documentation for the image-creation large model sits under the platform's AI and machine learning category. That is the path for anything you intend to run at volume, and it is the only path that gives you the rate card above.

Then there are the first-party products, and this is where the preview label understates how available the model already is. Tencent has put Hy Image3.5 preview into Yuanbao, its consumer assistant; into the ima knowledge tool; and into the creative and productivity surfaces Miora, WorkRally, OnSolo and WorkBuddy. It is also wired into Tencent Cloud's media stack through Media Processing Service and Video on Demand, which means the intended buyers include teams already transcoding and editing video on Tencent infrastructure rather than only developers building image features from scratch.

And then there is the third-party route, which changed shape recently. ComfyUI shipped partner nodes for the model — two of them, from version 0.37.1 — and the important detail is what kind of node they are. They call Tencent's API. There are no local weights, no downloadable checkpoint, no offline path. That is worth knowing before you plan around it, because "there is a ComfyUI node" and "I can run this on my own GPU" are different statements, and for this model only the first one is true.

The free window is the reason the naming matters right now

Miora, WorkRally and OnSolo opened a free two-week window on 22 September 2026 that closes on 7 October 2026. That window is the cheapest way anyone outside Tencent has had to form an opinion of this model, and it is attached to the products, not to the API — so a team that wants to evaluate it before committing to a pipeline should be reading "free on Miora, WorkRally and OnSolo" and not "free on Tencent Cloud." The API is metered from the first call.

That distinction is the one place where the two names genuinely lead you to different places. Search the short name and you will find the free window. Search the long name and you will find the rate card. Both are correct, and a team that only read one of them would have got half the picture — which is a fair summary of why this piece exists at all.

What the evidence does and does not support, under either name

The vendor claim is a GSD-style blind preference test run with what the launch materials describe as several hundred of Tencent's own professional designers, reporting roughly a thirty percent improvement over Hy Image3.0, parity with ByteDance's Seedream 5.0 Pro, and a marginal edge over Nano-Banana Pro and Alibaba's Qwen-Image-3.0-Pro. Vendor-run, vendor-populated, prompt set unpublished, unreproduced. The methodology is a reasonable one and the participants are Tencent employees; both things are true at once, and the honest label is vendor-reported.

The independent board has nothing under either name. Artificial Analysis's text-to-image board, read on 23 September 2026, carries no Hy Image3.5 row and no Hy Image3.5 preview row, because there is only one model and it is not on the board. What it does carry from Tencent is the previous generation: HunyuanImage 3.0 Instruct at 964 Elo and HunyuanImage 3.0 at 945, ranked 65th and 70th of 156 models. For scale, the models Tencent claims to have matched sit far above — Seedream 5.0 Pro at 1078, Qwen-Image-3.0-Pro at 1088, Grok Imagine Image 2.0 at 1154, and the leader, GPT Image 2.5 Sunburst (max), at 1197.

There is one number in circulation that fails against Tencent's own announcement, and it is worth flagging because it spreads under the short name. At least one outlet reported a 4096×4096 output spec. Tencent says "up to 2K" in as many words, and every other source agrees. Treat the 4K figure as an error, not as a roadmap for the official release.

Where this leaves a team deciding today

If you are building something that has to survive a vendor's release calendar, the naming question is the answer to your real question and the answer is uncomfortable: there is no stable identifier to pin to. The model is a preview, the only identifier is a preview identifier, and Tencent has not published a date for the official version. What you can do is write the generation call so that the model name is a configuration value rather than a constant, which costs nothing and is the same discipline that lets you move between a $0.024 model and a $0.21 one without a sprint.

Where that discipline pays off is at the routing layer. OrcaRouter runs on a 0% markup model — the provider's list price passed through unchanged, so a vendor price revision lands on our side the same day rather than waiting on a repricing cycle — and one API key covers 200-plus models, with automatic failover across providers and a routing DSL for pinning a specific model per request. To be exact about our own catalogue: we do not currently route Hy Image3.5 preview under either name, so the model itself means Tencent's own cloud and the first-party products listed above. What the routing layer gives you is somewhere stable to stand while you wait for the preview to become a release — openai/gpt-image-2, google/gemini-3.1-flash-image-preview, the Imagen 4 family and grok/grok-imagine-image all sit behind that same endpoint, which is what makes the wait productive rather than idle.

The short answer, if you only wanted the short answer

They are the same model. The long name is the accurate one and the short name is what everyone calls it. The identifier you will need is hy-image-v3.5-preview, the international rate is $0.024 per 2K image billed on delivered output, the free window closes 7 October 2026, and no independent leaderboard has measured it yet under either name. The next thing that will change this picture is either a stable release with a date attached, or a first third-party Elo — and the second of those is the one that would tell you whether the thirty percent is real.