Tencent rolls out Hy Image 3.5 preview

Tencent Holdings has introduced Hy Image 3.5 Preview, its newest image-generation model, as it intensifies its push into generative artificial intelligence amid growing competition from entities such as ByteDance and Alibaba Group. The model facilitates text-to-image creation, image-to-image generation and multi-round editing, as outlined in Tencent Cloud documentation accompanying the release.

It can process multimodal inputs combining text and reference images, maintain context across successive editing turns, and produce images with resolutions reaching up to 4K. Tencent claims the model enhances Hy Image 3.0 in terms of text rendering, realism, stylistic expression, and editing consistency, demonstrating around a 30 percent overall capability improvement in internal evaluations, though these claims have not been independently verified.

The new system is integrated into Tencent's products, including Yuanbao chatbot and various design, productivity, and creative tools, with developer access available via Tencent Cloud’s TokenHub platform. The model is identified as hy-image-v3.5-preview on Tencent’s cloud documentation, supporting up to 100,000 input tokens for conversational image-generation and editing tasks.

Tencent emphasizes iterative visual production over single-prompt image creation, enabling users to provide reference pictures, request edits across multiple rounds, and retain earlier context for tasks like poster design, product-image editing, and other workflows requiring repeated adjustments. The model can utilize external search enhancement and automatic prompt rewriting, allowing developers to specify image dimensions or let the system determine an appropriate format.

It supports output up to 4,096 by 4,096 pixels under Tencent Cloud’s documented limits. This launch follows China's major tech companies expanding their AI offerings beyond large language models, intensifying competition in image, video, and agent-based systems. ByteDance leads in visual-generation tools through its Seedream family, while Alibaba has expanded its Qwen image models alongside a broader AI infrastructure and chip push.

Tencent claims Hy Image 3.5 Preview performs comparably to ByteDance’s Seedream 5.0 Pro and slightly outperforms Google’s Nano Banana Pro and Alibaba’s Qwen-Image-3.0 Pro in internal assessments, though full methodologies and independent benchmark results for these comparisons are not disclosed. The timing aligns with Alibaba’s Apsara conference in Hangzhou, where the company announced plans for a larger next-generation AI model and introduced Zhenwu V900 accelerator chip.

Alibaba CEO Eddie Wu stated the planned model would have parameters ranging from five trillion to 10 trillion, potentially outpacing current flagship systems. Tencent has been intensifying investment in its Hunyuan model family, focusing on deployment across existing products. Its Hy3 language model, released in July, was designed using a mixture-of-experts architecture and integrated into services like WorkBuddy, Yuanbao, and Tencent Cloud.

Tencent’s former OpenAI researcher Yao Shunyu serves as chief AI scientist, while computer-vision specialist Tian Yonglong leads vision-language model development within the Hunyuan team. Tencent’s Hong Kong-listed shares rose more than 6 percent during Tuesday’s trading as investors reacted to the product launch and broader technology stock gains.