On September 22, 2026, Tencent's Hunyuan team announced Hy Image 3.5 preview (Hunyuan Image 3.5 preview), a new AI model for generating and editing images. It is designed for use in professional production settings and supports text-to-image generation, image-to-image generation based on an existing image, and editing that revises images over a multi-turn conversation. As the word "preview" in its name indicates, it is positioned as an early release ahead of the official version.
The announcement
Hy Image 3.5 preview is the successor to the previous-generation Hy Image 3.0. According to Tencent's official announcement, multiple reference images can be passed in a single request, and it can output in a variety of aspect ratios. It also supports "multi-turn editing," in which the same image is revised little by little while continuing the conversation, taking into account the earlier exchanges.
With earlier image generation AI, when you wanted to partially fix an image you had already created, there were many cases where you had to rewrite the instructions from scratch. With multi-turn editing, you can layer additional instructions such as "make only the background night" or "make the text a little bigger," allowing a way of working close to revising a design.
Number of reference images and resolution
The number of reference images and the output resolution are described differently depending on the document. The official announcement on September 22 stated that up to five reference images can be used at a time and that output is up to 2K.
On the other hand, Tencent Cloud's API documentation for developers lists, as specifications for the model ID hy-image-v3.5-preview, up to 20 reference images in total across a conversation (separate from the per-request limit) and output of up to 4096×4096 pixels (4K). Input can be up to 100,000 tokens (the unit in which AI processes text), a large expansion from the previous-generation Hy Image 3.0 (up to three reference images and output with an area equivalent to at most 1024×1024 pixels). It also offers a feature in which the AI rewrites the prompt (the instruction text) while thinking it through, a feature that draws with reference to external search results, and a feature that adds watermark text to images.
The number of images and resolution available may differ between using it from apps and using it via the API. It is a good idea to check each interface and its documentation before use.
Tencent's own evaluation
Tencent says hundreds of professional designers within the company carried out a comparative evaluation called "GSB," in which they compared images to judge which was better. As a result, it explains, Hy Image 3.5 preview improved 30% over the previous-generation Hy Image 3.0, was on par with ByteDance's Seedream 5.0 Pro and slightly outperformed Google's Nano Banana Pro and Alibaba's Qwen-Image-3.0 Pro. All of these are Tencent's own evaluations, and no external verification has been presented.
Tencent also says that when it gradually trial-ran the model in its AI assistant Yuanbao starting about two months before the release, the number of image generation requests rose by more than 50% overall.
Apps where it is available and API pricing
Hy Image 3.5 preview has been built into the following Tencent apps and services.
- Yuanbao, an AI assistant
- WorkRally, for video production
- OnSolo, for creating drama and game content
- Miora, an AI agent for creative support
- WorkBuddy, an AI agent for work
- ima, an AI knowledge management tool
- Tencent Cloud's video processing services (MPS and VOD)
Tencent says features such as portrait photo editing can be used for free in Yuanbao, and Miora, WorkRally and OnSolo can be tried for free for a limited period of two weeks.
For developers, it is offered through the Tencent Cloud API. The price is 0.15 yuan per 2K image, and $0.024 per image on the international version for overseas users. Only output images are charged, and inputs such as reference images are free. Tencent has provided no guidance about availability in Japan or use in Japanese.
Stock price reaction and the competitive landscape
According to Bloomberg, Tencent's shares rose more than 7% in the Hong Kong market following the announcement. Investing.com reported that as of around noon Japan time, they were up 6.6% at HK$458.4.
The announcement coincided with the day Alibaba held its annual event Apsara Conference (Yunqi Conference). Bloomberg reported that Tencent is trying to catch up with ByteDance and Alibaba in image generation AI.
CoAI's take
What stands out about Hy Image 3.5 preview is less the model's performance itself than how widely it has been built into Tencent's many products, from everyday apps such as Yuanbao and ima to video and game production tools. Its price of $0.024 per 2K image is also about half the level of Google's Nano Banana 2.1 (about $0.050 for a 2K image). On the other hand, the performance comparisons are only Tencent's own evaluations, and the number of reference images and resolution are also described differently across documents. Competition in Chinese image generation AI is intensifying with ByteDance's Seedream and Alibaba's Qwen-Image, and when using it from Japan, it is best to try it after checking the regions where the API is offered and the terms of use.
Sources
- Hy Image3.5 preview发布！腾讯多产品限时免费体验 (Tencent News)
- 大模型服务平台 TokenHub Hy 生图调用指南 (Tencent Cloud)
- 腾讯混元 Hy Image3.5 preview 发布：支持文生图与多轮对话创作，已上线元宝、ima 等应用 (IT Home)
- Tencent Releases AI Image Model to Catch ByteDance, Alibaba (Bloomberg)
- Tencent shares jump after unveiling new AI image model (Investing.com)
- Gemini Developer API pricing (Google AI for Developers)