news
Tencent ships Hy Image 3.5 Preview at $0.024 an image, with a 30% win rate claim it scored itself
The Hunyuan team's new image model merges text-to-image, reference input and multi-turn editing, but the only quality number attached to it comes from Tencent's own human evaluation.
Tencent released Hy Image 3.5 Preview on September 22, 2026, and the headline quality figure attached to it is a 30% win rate over its own previous model in a human evaluation Tencent ran on itself. The launch post does not say how many evaluators took part, what prompts they saw, how ties were handled, or which competing models were in the comparison pool. No independent benchmark had a score for the model on launch day.
One model covering generation, reference input and multi-turn editing
The practical change is in how the model is called rather than in any single output. Hy Image 3.5 Preview handles text-to-image and image-to-image in the same model, accepts reference images, and carries prior images and conversation history forward so an application can request a targeted revision instead of regenerating from scratch. Tencent Cloud lists it under the identifier hy-image-v3.5-preview. That moves image generation closer to how editing tools actually get used, where the second and third attempt matter more than the first.
Pricing in the announcement is $0.024 per generated image on the Tencent Cloud API, with reference images not charged. Tencent also said two of its own surfaces, Miora and OnSolo, would offer free access for two weeks. Beyond that, the model is being folded into Tencent products including the Yuanbao chatbot and its design and film-editing tools.
The resolution ceiling depends on which Tencent source you read
Tencent's English-language launch post describes output at up to 2K. Chinese-language coverage of the Miora rollout describes direct generation at 4096x4096 and a reference-image limit raised from three to twenty, and the Tencent Cloud documentation for the model covers explicitly requested 4096x4096 output. Tencent has not published a single specification page that reconciles those figures, so the safe reading is that 2K is what the global launch material commits to and 4K is what some surfaces expose.
A win rate without a published protocol is a marketing number
A 30% win rate is not meaningless, but without the prompt set, the evaluator pool and the tie-handling rule, it cannot be reproduced or compared against anyone else's 30%. This is the same gap that shows up across the image and video leaderboard landscape, where rankings get published without prompt sets, judge identity or vote counts. Tencent also published no parameter count, no architecture description, no technical report and no stated plan for weights, which is a step back from Hunyuan Image 3.0, whose base was released openly.
The timing was aimed at a competitor's stage
The release landed on the same day Alibaba opened its Apsara Conference in Hangzhou, and Bloomberg framed it as Tencent signaling intent to close the gap with ByteDance and Alibaba in image generation. For anyone actually deciding whether to route work to it, the useful facts are the per-image price, the free reference inputs and the multi-turn editing path. The 30% is a claim to test, not a result to cite.
Sources: Bloomberg, Tencent Hunyuan.
Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.