NExT-GPT

Any-to-any multimodal LLM (legacy anchor, 2023): one of the first open systems accepting and producing arbitrary combinations of text, image, video, and audio, by connecting an LLM core with multimodal adaptors and diffusion decoders — trained with only ~1% of parameters updated via modality-switching instruction tuning (3.6K★).

Pre-dates this index's main window but is the load-bearing precedent for the NExT++ lab's multimodal-generalist agenda (Tat-Seng Chua, Hao Fei) that leads to General-Level & General-Bench and Vitron. Built on open LLM/diffusion components, not a from-scratch pretrain.