“Qtai's Moshi is a 7 billion parameter model featuring a dual-channel architecture with a joint attention mechanism between its audio input and output channels, allowing it to decide when to interrupt.”