Uni-MoE is a family of large multimodal models based on Mixture of Experts architecture. It supports understanding and generation across text, image, audio, and other modalities. The project provides multiple model versions with corresponding training code, pretrained checkpoints, and academic publications released through Hugging Face.