Tencent: Hy Image 3.5 Preview

Tencent: Hy Image 3.5 Preview is an image generation and editing model from Tencent, released as a preview on 5 October 2026. It sits in the image category and is built as a unified model, meaning generation and editing are handled by the same endpoint rather than separate specialised systems. If you are evaluating image models for a product or internal tool, this is Tencent's current preview offering in that space.

The model accepts text and image inputs and produces image outputs. It supports text-to-image, image-to-image, and multi-turn editing workflows through one endpoint, and it can take up to 20 reference images in a single request. The context window is 100,000 tokens, with a maximum output of 90,000. Image output is configurable through several parameters: resolution (1K, 1.5K, 2K, or 4K), aspect ratio (a choice of 13 ratios including 1:1, 3:2, 4:3, 16:9, 9:16, and 4:1), and a seed option. The n parameter is fixed at 1, so each request returns a single image.

Those capabilities point to a few practical uses. Text-to-image covers straightforward asset creation from a written prompt. Image-to-image and multi-turn editing suit iterative work, where you refine a result across several turns instead of rewriting the prompt from scratch each time. The 20 reference images are useful when output needs to stay consistent with existing material, such as matching a product, a character, or a visual style across a set of images. The resolution and aspect ratio options let you target specific placements, from square social formats to wide or tall layouts, without post-processing crops.

You call the model through api14's OpenAI-compatible API using the model id "tencent/hy-image-v3.5-preview". Because the interface follows the OpenAI format, you can point an existing OpenAI client at api14 and change the model id. The model itself lists no supported parameters in the standard request schema, so the controls you use are the image parameters: resolution, aspect_ratio, n (fixed at 1), input_references (0 to 20), seed, and size. Plan your requests around those fields rather than assuming sampling controls like temperature or top_p are available.

One limit worth noting is that n is capped at 1, so you cannot request multiple variations in a single call; if you need several options, issue separate requests, optionally varying the seed. The model is also a preview release, which is worth weighing if your workload depends on long term stability. If your task is text generation or another non-image modality, a different model on api14 will fit better, since this one outputs images only.