Tencent's HunyuanImage-3.0 Challenges Diffusion Models With 80B Open Autoregressive Giant
Tencent's 80B parameter Mixture-of-Experts model ditches diffusion transformers for a unified autoregressive framework that reasons before it renders.
Takeaways
- Tencent open-sources HunyuanImage-3.0, an 80B parameter MoE with 13B active per token.
- Uses a unified autoregressive framework with Transfusion-style diffusion, not a DiT backbone.
- Trained on 5B image-text pairs and 6T text tokens using the Hunyuan-A13B foundation.
- Instruct variant adds Chain-of-Thought reasoning, image editing, and multi-image fusion up to 3 inputs.
- Distilled checkpoint runs in 8 sampling steps; vLLM and FlashInfer supported for faster inference.
- Released under Tencent Hunyuan Community License permitting commercial use; technical report available.
Tencent has open sourced HunyuanImage-3.0, and the headline number is worth pausing on: 80 billion total parameters with 13 billion active, arranged as a Mixture of Experts across 64 experts. That makes it the largest open-weights text-to-image model shipped to date. The more interesting story is architectural. Instead of following the industry into ever-larger diffusion transformers, Tencent built an autoregressive multimodal model that treats image generation the way an LLM treats the next token.
Understanding and generation live in the same autoregressive framework, and the image half is what Tencent has released publicly. The system leans on careful data curation, a native Chain-of-Thought schema, progressive pre-training, aggressive post-training, and infrastructure tuned for large-scale MoE inference. Weights are available on Hugging Face and GitHub under the Tencent Hunyuan Community License, which permits commercial use.
An autoregressive bet against diffusion orthodoxy
The core design choice breaks with almost every other frontier image model. Rather than a Diffusion Transformer (DiT), HunyuanImage-3.0 runs on Hunyuan-A13B, a decoder-only MoE transformer that also powers Tencent's language models.

Diffusion isn't gone, though. It's spliced into the LLM stack in the same manner as Transfusion, which the paper cites as the direct architectural precedent. In practice, the model predicts text tokens autoregressively and diffuses image tokens inside the same backbone, so understanding and generation share weights instead of being bolted together with adapters.