Ming-Lite-Omni-Preview: A MoE Model Designed to Perceive a Wide Range of Modalities
MoE model handles text, images, audio, and video with streaming speech generation.
New model
Ant Ling released Ming-Lite-Omni-Preview, a Mixture-of-Experts model built on Ling-Lite that processes text, images, audio, and video while generating text and natural speech in a streaming manner.
The model incorporates modality-specific routers and supports video understanding with KV-Cache dynamic compression, natural speech generation with dialect support and one-shot voice cloning, and a chain-of-thought reasoning mode.
It is available for download from 
Hugging Face and ModelScope.
