Tencent Timed Its Image Model Launch to Alibaba's Keynote. That Is the Whole Story.

On September 22, Tencent put Hunyuan Image 3.5 Preview online. The same day, Alibaba opened its Apsara conference, the company's flagship annual AI event. The overlap was not a coincidence, and Chinese tech media treated it as exactly what it was: a marketing move dressed as a product launch.

This is worth paying attention to because it tells you what the image model race in China has become. The models are getting better, but the more revealing change is how the players now fight for attention. Launch day has turned into a weapon.

What Tencent Shipped

Hunyuan Image 3.5 Preview is an image generation and editing model. Compared with the previous Hy Image 3.0, Tencent says it improved text rendering, layout, realism, and style handling, and added continuous editing. It supports both text-to-image and image-to-image, and you can feed it up to five reference images at once for a single generation, with output up to 2K resolution.

The practical change is that image creation is no longer a one-shot prompt. You can generate a poster, then keep editing it in the same conversation, changing text, swapping elements, and adjusting the layout, without starting over each time. That continuous, multi-turn editing is the same direction the rest of the field has been moving in all year.

Tencent's numbers come from internal testing. The company says a blind test with several hundred professional designers put Hy Image 3.5 Preview about 30 percent ahead of Hy Image 3.0 in overall win rate. It says the model is roughly level with Seedream 5.0 Pro and slightly ahead of Nano Banana Pro and Qwen-Image-3.0 Pro. Those comparisons are Tencent's own, and no independent evaluation has confirmed them yet, which is worth keeping in mind.

Pricing is aggressive. On the Tencent Cloud API, a 2K image is 0.15 yuan, and the company only charges for the final output image. The model already runs inside several Tencent products, including Yuanbao, WorkRally, OnSolo, Miora, WorkBuddy, and ima. Free users can access portrait editing, smart retouching, travel planning images, and knowledge infographics in Yuanbao, while businesses and developers go through the cloud API.

Why the Date Mattered

The launch landed on the exact day Apsara opened. The logic is simple and effective. When Alibaba gathers the press, developers, and analysts for its biggest event of the year, the country's tech attention is already pointed at one place. Releasing your own model that day means catching a share of that attention without spending on a separate media push. The market noticed. Tencent's stock moved sharply around the announcement, according to MarketWatch.

This is not a one-off trick. In China, the announcement date has become its own competitive move. Earlier in the year, GLM-5.3-Flash and Qwen3.8-Flash-Next launched on the same day, a sign that synchronized launches have become a reflex. The pattern is spreading because it works.

Alibaba did not show up empty-handed. At Apsara, Eddie Wu talked about a future flagship model with 5 to 10 trillion parameters and an in-house accelerator chip, the Zhenwu V900. The company is playing on language models, images, and silicon at the same time. One more detail that says a lot about intent: Tian Yonglong, a researcher who worked on vision at OpenAI, joined the Hunyuan team back in July, weeks before a major image launch. Hiring a vision specialist trained at a rival right before you ship is not a routine HR note.

The Crowded Field Behind the Timing

The launch-day tactic only works because there is a real race to have a launch. Image generation has become one of the hottest fronts in China's AI competition, with ByteDance pushing Seedream, Alibaba rolling out the Qwen-Image line, and Tencent making clear it will not watch from the sidelines. In that setting, the players are no longer content to simply answer each other's releases. They deliberately schedule their own to collide with rivals' events and steal the spotlight. In a single day in late September, the market got both of the competing models at once, a closed model at a bargain price from Tencent and an open-weight release with a restrictive license from Alibaba.

That collision is not a coincidence of calendars. It reflects a market where the underlying technology is converging and differentiation has to come from elsewhere. When two companies can both produce a poster with legible text, change a background without breaking the subject, and hold a style across edits, the models start to look interchangeable in a demo. The fight then moves to price, to license terms, and to who owns the moment the market is paying attention. Tencent choosing Apsara's opening day is a bet that the last of those three matters more than it used to.

Two Strategies, One Day

What makes the September 22 pairing useful is that it showed two competing strategies side by side.

Tencent went with a closed model at a low price. You cannot download it, you rent it through the API, and the price is set to pull in product teams, e-commerce sellers, and editing pipelines that care about cost per image more than they care about owning the weights.

Alibaba, on the image side, went the other way. It released Qwen-Image-2.1 as open weights, a 7B model, along with two 9B prompt-rewriting checkpoints. But the license moved away from the permissive Apache 2.0 toward a strictly non-commercial research license. So you can download it, study it, and build on it for research, but commercial use runs into a wall. That is a different trade: maximum reach and mindshare, with the monetization question left open.

Read together, the two moves describe the fork the whole field is facing. Charge little for access and keep control, or give away the weights and keep the users, but not the revenue. Nobody has settled on which answer wins.

What It Means If You Build With These Tools

For a team picking an image model, the launch-day theater is noise and the licensing terms are signal. A cheap rate per image can change your unit economics overnight, the way a 0.15 yuan 2K price makes certain high-volume workflows viable. A non-commercial license can just as quickly rule a model out of a commercial product, no matter how good the demos look.

The other signal is the multi-turn editing. Every serious model is converging on the same idea: generation is the easy part, and iterating on a single image without breaking what already works is where the value sits. Tencent is charging for the final image only, which quietly encourages you to iterate inside the tool. That is a bet that the editing loop, not the first render, is where users will spend their time.

The image model war in China is now running on two tracks at once. On one, companies race to make the most capable model. On the other, they race to make it cheap and to own the moment it is announced. September 22 was a clean example of both tracks meeting on a single day, in a single news cycle. The model that wins the crowd that day is not necessarily the model that wins the year, but in a market moving this fast, being seen on the right day is starting to count for almost as much as being good.