HunyuanImage-3.0

At 80 billion total parameters with 13 billion activated per token, HunyuanImage-3.0 is the largest open-weights text-to-image model released to date. Tencent built it as a native multimodal Mixture-of-Experts model with 64 experts that unifies understanding and generation in a single autoregressive framework rather than the usual DiT plus text encoder split, which lets it perform reasoning-based prompt expansion, image editing, and multi-image fusion alongside straight generation. The hardware bill is steep: the README calls for at least three 80 GB GPUs to run the base model and eight for the Instruct variants, on CUDA 12.8 with PyTorch 2.8, though a distilled Instruct release and vLLM acceleration support arrived in January 2026 to ease deployment. Weights are on Hugging Face under the Tencent Hunyuan Community License, which allows commercial use for services below 100 million monthly active users but excludes the EU, UK, and South Korea from its territory. Open-sourced in September 2025 with a technical report on arXiv, the repository has around 3,200 stars.