The Pulse
Tencent Previews Hy-Image 3.5 With 4K Multi-Turn Editing
Tencent Cloud has documented Hy-Image-3.5-Preview, a new image model that supports text-to-image, image-to-image generation and multi-turn editing. The service can produce images up to 4096×4096 pixels and accept as many as 20 reference ima
AI.info Team ·
Tencent documents a larger Hunyuan image model
Tencent Cloud has published documentation for Hy-Image-3.5-Preview, a Hunyuan image-generation model built for text-to-image creation, image-to-image generation and multi-turn editing. The documentation was updated on September 21, 2026, and describes the model as a preview service available through Tencent's TokenHub platform.
Unlike Hy-Image-3.0, which uses a conventional prompt-and-size request, the new model accepts an OpenAI-style messages conversation. That structure allows an application to send an initial prompt, receive an image, then submit a follow-up instruction with the previous conversation history for another edit. Tencent says the service returns the result synchronously rather than requiring a separate job submission and polling cycle.
The model is identified in the API as hy-image-v3.5-preview. Its documented endpoint is a Tencent-hosted POST interface at https://tokenhub.tencentmaas.com/v1/wand/hunyuan-image/v35-generation. Access requires a Tencent Cloud account, TokenHub activation and an API key.
4K output requires an explicit size
Hy-Image-3.5-Preview supports output dimensions from 256 pixels to 8,192 pixels on each side, subject to a total area limit of 16,777,216 pixels. That cap allows a maximum square output of 4096×4096, which Tencent labels as 4K.
Users must pass a specific size value to request the maximum resolution. Tencent's example uses 4096x4096. If the parameter is omitted, the service chooses dimensions based on the prompt and a separate pixel-budget setting. Tencent's documentation says that automatic pixel-budget modes reach 2K, while direct 4K generation requires the explicit size parameter.
The distinction matters for applications that need predictable production assets. A design tool can force a fixed canvas, while a conversational workflow can leave the dimensions to the model when speed or flexible composition matters more than a predetermined output size.
Conversation history becomes the editing interface
The multi-turn workflow keeps the original user message, the model's response and generated-image records in the next request. Tencent supplies an assembled_history structure in the response so developers can append the returned context without reconstructing the service's internal streaming messages.
For the next round, the application adds a new user instruction to the prior messages. The final user message becomes the active image-generation command, while the earlier messages provide context for edits such as changing a subject, adjusting a style or revising a composition.
Tencent also documents support for text and reference images in the same message. Hy-Image-3.5-Preview accepts public image URLs or Base64-encoded images, with each image limited to 20MB and the total image count limited to 20. The reference-image table lists PNG, JPEG and JPG support.
More input, more controls than Hy-Image-3.0
The preview model accepts up to 100,000 tokens of input, compared with an 8,192-character prompt limit listed for Hy-Image-3.0. Tencent's documentation also lists external-search enhancement, prompt or reasoning rewrite, custom dimensions, area controls, random seeds and image watermark footnotes among the available capabilities.
Developers can use a seed between zero and 263-1. Tencent says a fixed seed, prompt and size are required when an application needs to reproduce the same result. The service can also select an output size from the prompt when no exact dimensions are supplied.
Both Hy-Image versions operate synchronously through TokenHub. Tencent says Hy-Image-3.5-Preview performs its internal SSE-based reasoning on the server and returns the final image frame when processing reaches a terminal state. The application does not need to assemble incremental SSE output itself.
Temporary image URLs put storage responsibility on developers
The generated-image URLs are temporary and expire after 12 hours. Tencent directs developers to download and store the final image promptly. For multi-turn editing, applications also need to preserve the intermediate image URLs contained in the assembled history if they intend to feed that context into another request.
Tencent's documentation positions Hy-Image-3.5-Preview for professional visual work such as posters, user-interface design and product-image editing. The practical change is less about a new prompt format than about treating image creation as an ongoing conversation: reference images, prior outputs and new instructions can remain in one request history, while a precise size parameter can push the final image to 4K.
Tencent Cloud's Hy image-generation guide provides the request schemas, examples and output-field details for both Hy-Image-3.0 and Hy-Image-3.5-Preview.