Hy Image 3.5 Preview: what Tencent shipped
Tencent's Hy Image 3.5 Preview generates and edits in one model; its API guide lists 4K output and 20 references. No weights or report. Rows to use today.
Tencent released Hy Image 3.5 Preview on 22 September 2026 (Cellcog). It is the successor to Hunyuan Image 3.0, under Tencent's shorter "Hy" brand. The pitch is one model for three jobs: text-to-image, reference-based generation and multi-turn editing. You no longer route a generate to one model and the fix to another.
It is a preview, and a closed one. There are no weights, no parameter count and no technical report. Versely does not serve it. Below: what is confirmed, what conflicts, and which rows to use for the same jobs today.
What is confirmed
One model, three jobs. Per Cellcog's launch write-up, "one model handles text-to-image, reference-based generation and multi-turn editing". That matters in practice. When the same model makes and fixes the image, the fix does not drift from the look of the first render.
Resolution and references, with a caveat. The two launch documents disagree. The announcement says up to 2K output and a maximum of 5 reference images. The API guide lists sizes "up to 4096 by 4096" and up to 20 input images at 20 MB each. Our reading: the API is the more capable surface, and the consumer apps may be capped lower. Test the endpoint you will actually call before you build a 20-reference workflow around it.
Price. Per Cellcog, Tencent's international API charges $0.024 per image at the 2K tier. Mainland pricing is 0.15 RMB per image at 1K and 2K, and 0.20 RMB at 4K. Billing counts output images only. Reference inputs and failed generations are free. That last detail is rare and useful: a failed call costs nothing.
Why one model for both jobs matters. Most image pipelines today are two models glued together: a generator for the first render and an editor for the fixes. Every hand-off is a chance for the look to drift. The lighting shifts, the skin tone moves, a logo gets redrawn. A single model that holds the scene across rounds removes that seam. Tencent is not the first to try this. GPT Image 2.5 and Nano Banana Pro both have generate and edit rows on the same model. It is the direction the whole category is moving, and Hy Image 3.5 is Tencent's entry.
Where it runs. Inside Tencent's own apps (Yuanbao, WorkRally, OnSolo, Miora, WorkBuddy, ima) and on Tencent Cloud MPS and VOD.
What is not known
No weights and no report. Cellcog notes "no parameter count, no architecture, no technical report, and no open-weights plan". For context, Hunyuan Image 3.0 shipped in September 2025 as an 80B mixture-of-experts model with 13B active parameters. Do not assume 3.5 is built the same way, and do not plan to self-host it.
No independent benchmark. The only quality number is Tencent's own: a blind test by several hundred Tencent designers, reported by Cellcog as a "30% win rate over Hy Image 3.0". No public leaderboard had scored it on launch day. Until one does, treat quality claims as marketing.
Text rendering and language support. Not specified in the launch material. If your brief has a headline in the image, test that before anything else.
Versely's Hunyuan rows today
Versely does not serve Hy Image 3.5. It serves the previous generation as two rows:
- HunyuanImage 3.0 Instruct Text to Image, 8 credits. Prompt in, image out. No references.
- HunyuanImage 3.0 Instruct Edit, 8 credits. A still plus a plain-language instruction. Up to 3 references. It keeps the structure of the source and changes what you asked for.
So you already have Hunyuan's generate-then-edit loop. It is two rows, not one model, and it takes 3 references, not 20. Hunyuan Image 3.0 Instruct Edit is an instruction, not a generate covers the edit row's rules. The short version: give it a still and one clear change per call.
Hy Image 3.5 against the rows that win today
A preview with no public scores is not something to build on yet. The useful question is what each image model is for, and where 3.5 would slot in if it reached the catalog.
| Job | Best row today | Why |
|---|---|---|
| Precise, repeated edits to one still | GPT Image 2.5 Sunburst Edit | Edits scoped to the instruction, subject kept across many rounds, optional mask, up to 16 inputs |
| Hero still with fine detail | GPT Image 2.5 Sunburst Text to Image | Precision tier, up to 4K, transparent background |
| Reference-heavy consistency (a product, a face, a set) | Nano Banana Pro | Up to 8 references, 4K, generate and edit |
| Generate-then-fix loop in one model family | HunyuanImage 3.0 Instruct pair | Same model family for the render and the fix |
| Many references in one call (10+) | Hy Image 3.5, if the API's 20-input limit holds | Nothing in the catalog takes 20 references today |
Where GPT Image 2.5 Sunburst wins. Sunburst is OpenAI's precision tier from the ChatGPT Images 2.5 launch on 8 September (9to5Mac). Its edit row takes an optional mask, which Hy Image 3.5's launch material does not mention. A mask is the difference between "change the label" and "change the label, and please do not touch the hand holding the bottle". Sunburst runs 4 to 7 credits on Versely depending on quality tier. For the tier choice, see GPT-Image-2.5 Flare vs Sunburst.
Where Nano Banana Pro wins. Reference discipline. Eight references at 4K, generate and edit rows on the same model: this is the pick when the product has to look like the product in every frame. It is 12 to 24 credits for a generate and 6 to 12 for the edit row. It costs more than Sunburst per image. You pay for consistency.
Where Hy Image 3.5 might win. Two places, both unproven. First, raw reference count: 20 inputs, if the API guide is right, beats both 8 and 16. Second, API cost for anyone buying from Tencent directly: $0.024 per 2K image with free failures is aggressive. Neither matters until independent tests show the output holds up. More references are not better if they fight each other. Reference image hygiene explains why 20 badly matched references lose to four clean ones.
Who should care
Teams already on Tencent Cloud. Hy Image 3.5 is in MPS and VOD. Test it there first.
Anyone building a reference-heavy pipeline. Catalogue shots, character sheets, a room from several angles. The 20-input limit is the reason to watch this model. Test with 5, then 10, then 20, and see where consistency breaks.
Not yet: anyone who needs open weights, a paper to cite, or a third-party score before a procurement sign-off. None of those exist.
What to do now
- Keep your current image rows. Use Sunburst for precise edits, Nano Banana Pro for reference consistency, and the Hunyuan 3.0 pair when you want one family for the render and the fix.
- If you test Hy Image 3.5 directly, compare it against Sunburst Edit on your worst real edit, not on a demo prompt. The first test should include a masked-area problem and a line of text.
- Watch for three things: an independent leaderboard score, a technical report, and a clear answer on 2K/5 references versus 4K/20. Until those land, 3.5 is a promising preview, not a production default.
To see every image row side by side, start from compare.