Tencent open-sources HunyuanImage 3.0

Background

Image generation models turn a text prompt into a picture, and the result depends on how well the model follows the prompt and how good the image looks. Many image generation models are built on diffusion transformers. Tencent's HunyuanImage 3.0 took a different route: it is a native multimodal model that handles understanding and generation in one autoregressive framework.^[1] In this context, multimodal means one model both reads images and text and produces images. Tencent published the model as open source, which lets outside researchers and developers inspect and use it.

What happened

On 28 September 2025 Tencent released the code, weights and technical report for HunyuanImage 3.0.^[1] The model unifies multimodal understanding and image generation in one autoregressive framework.^[1] It is a mixture-of-experts (MoE) model with more than 80 billion parameters, of which 13 billion are active for each token.^[2]

Key details

Tencent called the model the largest open-source image-generation MoE model.^[1] The authors of the technical report said that automatic and human evaluations show the model rivals previous state-of-the-art models in text-image alignment and visual quality.^[2] Text-image alignment describes how closely a picture matches its prompt, and visual quality describes how good the image itself looks.

The technical report is hosted on arXiv and the code is in a GitHub repository, so anyone can read the paper and download the code.^[1][2]

Why it matters

The release gave open-source users a large image model built on a mixture-of-experts design, with the technical report published alongside the weights.^[2] A model of this size is demanding to run, so the release is most directly useful to researchers and developers with suitable hardware. Because the code, weights and report came out together, a reader can compare the paper's description with the released implementation.^[1][2] Tencent's follow-up Instruct version, which added image editing and reasoning, came on 26 January 2026.^[1]