On September 22, 2026 — the same day the Apsara Conference opened — Tencent released Hy Image 3.5 Preview, a new version of its Hunyuan image model. This was no routine iteration: Tencent made "multi-turn conversational editing" the centerpiece of the update, and directly named ByteDance's Seedream 5.0 Pro and Google's Nano Banana Pro as benchmarks. According to Dow Jones media outlets, Tencent's Hong Kong shares surged as much as 7.8% on launch day.
What the new model can do
Hy Image 3.5 Preview supports text-to-image, image-to-image, and multi-turn conversational editing. According to Tencent's launch post, the model can output images at up to 2K resolution. The bigger changes are for developers: the new API adopts a messages-style structure similar to multimodal chat interfaces, letting a single request carry text, reference images, and prior edit history at once — image editing is no longer a one-shot "prompt in, image out" affair but can iterate continuously like a chat. Tencent Cloud documentation shows a single request supports up to 20 reference images, with 100,000 tokens of input context, plus fixed random seeds, custom sizes, and watermarking. API pricing is $0.024 per image, 25% lower than Hy Image 3.0's list price.
How it stacks up, per Tencent
| Comparison | Tencent's claimed result |
|---|---|
| Hy Image 3.0 | 30% higher win rate in blind tests |
| ByteDance Seedream 5.0 Pro | Roughly on par |
| Google Nano Banana Pro | Slightly ahead |
| Alibaba Qwen-Image-3.0 Pro | Slightly ahead |
Take it with a grain of salt: all of these figures come from Tencent's own testing — several hundred in-house designers participated in GSB (good/same/bad) blind evaluations, with no third-party benchmark backing. The claim of being "on par with Seedream 5.0 Pro" can only be treated as the vendor's word until independent evaluations appear.
Shipping beats benchmarks
Rather than leaderboards, Tencent is emphasizing "getting it into products." Hy Image 3.5 Preview will be integrated into the Yuanbao AI assistant, the cinema-grade production platform WorkRally, the creative Agent Miora, the office Agent WorkBuddy, and the IMA knowledge-base assistant; the API is open through Tencent Cloud, and international users can try it via Miora and OnSolo. Usage on Miora is free until October 7.
This continues the thinking of Tencent's chief AI scientist Yao Shunyu: don't chase leaderboards, focus on product integration and solving real problems. Notably, both Yao and former OpenAI computer vision researcher Tian Yonglong, who joined the Hunyuan team in July, are OpenAI alumni — Tencent's visual model team is accelerating its "professionalization."
How to read it: the stock price has voted
On launch day, Tencent's Hong Kong shares (0700) rose as much as 7.8% intraday, closing up about 5.5% at HK$458.4 — the best single-day performance since early June. The logic the market saw is not complicated: while Alibaba told a long-term "10-trillion-parameter" story at the Apsara Conference, Tencent delivered a product checklist that can already be plugged into Yuanbao, WorkRally, and design workflows. Between a distant narrative and immediate shipping, capital markets this time voted for the latter.
