Tencent has released a preview of Hy Image3.5, its latest image generation model.

Hy Image3.5 is being integrated into Tencent’s consumer artificial intelligence, content creation, advertising, workplace, and cloud products. Teams behind those products are working with the Hy model team to improve its performance on specific tasks. Tencent calls the process co-design: products serve as testing grounds that help shape the model as well as channels for delivering it to users.

The approach is already at work across its Yuanbao AI assistant, WorkRally content creation platform, Miora creative agent, WorkBuddy office agent workspace, and Ima knowledge assistant. Hy Image3.5 is also available through Tencent Cloud, while international users can access it through Miora and OnSolo.

The model supports text-to-image and image-to-image generation and accepts up to five reference images. Tencent said its improvements over Hy Image3.0 focus on three areas: text rendering and layout, realism and stylistic expression, and consistency in image editing.

Tencent is also emphasizing cost. Hy Image3.5 is priced at USD 0.024 per 2K output image through Tencent Cloud, with customers charged only for images produced.

That pricing matters for the uses Tencent has in mind: commercial posters, television storyboards, promotional materials, user interface design, illustrations, brand visuals, and e-commerce assets, alongside everyday tasks such as photo retouching and creating educational materials.

These tasks depend on details that can be easy to overlook: legible small text, a character whose appearance remains consistent through repeated edits, or an image that can be reworked while preserving selected elements. Reliability in those details helps determine whether a model can produce hundreds or thousands of usable images without making the process prohibitively expensive.

Tencent said an internal blind evaluation involving hundreds of designers found that Hy Image3.5 performed on par with ByteDance’s Seedream 5.0 Pro and slightly better than Nano-Banana Pro and Qwen-Image-3.0 Pro. It also reported a 30% improvement over Hy Image3.0 in its own human evaluation.

Usage figures offer another measure of how the model performs in practice.

During a limited rollout in Yuanbao, Tencent said Hy Image3.5 increased overall image generation requests by more than 50%. Portrait retouching requests, which require precise editing, doubled, while requests for work tasks such as poster and logo creation rose 80%.

Other Tencent teams have helped shape the model around their requirements.

Tencent’s ima knowledge assistant integrated Hy Image3.5 a month before its release and worked with the Hy team to improve its performance on illustrations, posters, infographics, and image editing. Tencent said usage of ima’s image generation feature has since tripled.

WorkRally has tested the model with studios producing animated short dramas. Its product team said Hy Image3.5 can generate props, characters, and scene assets in batches while meeting most requirements for following instructions and keeping characters consistent. The Hy team has also worked with WorkRally on character stylization and making generated portraits look less artificial.

Tencent Ads said that, in its image-to-image applications, the incidence of problems such as garbled text and abnormal image redrawing fell 7% compared with Hy Image3.0.

Across these deployments, product requirements inform model development, and use of the resulting model generates feedback for the next round of improvements.

Discussion of foundation models often centers on advances in the models themselves: more parameters, better benchmark results, and stronger reasoning. With Hy Image3.5, Tencent is tying those efforts more closely to product development, making shortcomings encountered in daily use part of its research and development agenda.

Tencent has been reorganizing its model operations around this process. The company said that since an infrastructure overhaul began in February, the Hy team has released a major model update every two months on average. More recently, Tencent combined its large language model and multimodal teams into a foundation models department.

Preview releases allow researchers to incorporate feedback from business applications before a full release, the company said. Hy Image3.5 went through that process, including adjustments to align its visual output more closely with preferences observed in commercial use.

Tencent’s broad product portfolio gives it a range of settings in which to test and refine its models. The company is using that reach to bring the problems users encounter directly into model development.