China's Tencent released a new AI image-generation model called Hy Image 3.5 Preview between September 21 and 22, 2026. The model has been integrated into the company's Yuanbao chatbot and its Miora design tool, and Tencent says internal testing shows performance on par with ByteDance's leading model. The launch came on the same day Alibaba held its annual AI conference, intensifying competition among Chinese tech firms in AI image generation.

A Model Built for Multi-Turn Editing

Hy Image 3.5 Preview (API identifier: hy-image-v3.5-preview) generates images from text prompts and can also edit existing images based on instructions. It supports multi-reference generation using up to five input images, and multi-turn editing that preserves conversation history across revisions. Tencent also says the model has improved accuracy when rendering Japanese and English text inside generated images.

Tencent has built the model into several of its own products, including the Yuanbao chatbot, the Miora design tool, and video-editing tools, while also offering it to developers through the Tencent Cloud API.

A Few Cents Per Image, With Conflicting Resolution Specs

Pricing starts at $0.024 per image (about 4 yen) on the international API, and 0.15 yuan per image (about 3.5 yen) on Tencent's domestic China cloud service. Generating at 4K resolution on the China-only tier costs 0.20 yuan (about 4.7 yen).

※1 USD = 157 JPY, 1 CNY = 23.5 JPY (as of September 26, 2026)

The initial launch announcement described a maximum output of 2K resolution, but the developer-facing API documentation lists support for outputs up to 4,096 by 4,096 pixels (4K) — a discrepancy between the consumer-facing note and the technical documentation that has drawn attention on its own. It remains unclear how consistently the higher resolution is actually available. Tencent is offering free access to the model through its Miora design tool from September 22 through October 7.

Tencent Claims Parity With ByteDance and Google, But No Independent Verification Yet

Based on an internal evaluation involving hundreds of in-house designers, Tencent says the new model performs on par with ByteDance's Seedream 5.0 Pro, and slightly ahead of Google's Nano Banana Pro and Alibaba's Qwen-Image-3.0 Pro. The company also says the new model achieved a 30 percent higher win rate than its previous model, Hy Image 3.0, in human evaluations. All of these comparisons come from internal testing, and no independent, third-party benchmark results have been published so far.

The effort is being led by Chief AI Scientist Yao Shunyu, alongside Tian Yonglong, a former OpenAI computer vision researcher who joined Tencent in July. Tencent has faced criticism for lagging behind rivals in large language models, but the company has been hiring AI talent and shifting its focus toward shipping features inside its own products rather than chasing leaderboard rankings. Tencent's Hong Kong-listed shares rose more than 7 percent on the day of the announcement.

Summary

Tencent has released a new AI image-generation model, Hy Image 3.5 Preview, featuring multi-turn editing, multi-image reference support, and resolution up to a claimed 4K. Pricing is just a few yen per image, but the company's claims of matching or beating ByteDance and Google remain based on internal testing alone. Coming on the same day as Alibaba's major AI event, the launch underscores how intense the race among China's tech giants in AI image generation has become.