AnyGPT

Unified any-to-any multimodal LLM handling text, speech, image, and music via discrete sequence modeling — all modalities tokenized into a shared vocabulary, trained without architecture changes to the base LLM. A widely-cited early omni-modal LLM (ACL 2024) from Xipeng Qiu and Tao Gui's Fudan/OpenMOSS groups.