Block releases Dessa dialogue model trained solely on voice, which can run locally on MacBook According to official news from Block, Jack Dorsey's fintech company, Block's AI research lab, has released a research preview of its full duplex voice dialogue model, Dessa. This model is completely based on human dialogue speech trained from scratch, without using text or pre trained models, and can listen and speak simultaneously; Training only uses one set of NVIDIA GB200 NVL72, which can run locally at faster than real-time speeds on MacBook Pro M4 Max or higher configuration devices. At present, Dessa has not been fully opened. The official has released the dialogue demonstration recording and provided early access to a small number of AI security researchers. Project leader Ragavan Thurairatnam referred to it as the "GPT-3 moment in the field of speech".