According to CoinMeta, Jack Dorsey under Block has released a voice model called Dessa. What makes this model special is that it is trained from scratch entirely based on human dialogue recordings, without using any text data or relying on any pre-trained language models. Dessa directly receives and generates speech, supports full-duplex conversations, and can naturally respond before the other party finishes speaking. It can even laugh out loud when it hears something funny. Block mentioned that Dessa was trained using just one set of NVIDIA GB200 NVL72, which is equivalent to using 72 Blackwell GPU in total. The model's inference speed is faster than real-time playback, and it can run locally on devices equipped with M4 Max or higher-configured chips. Currently, Block has not disclosed the model's parameter size, standard test scores, or weights, and is only providing early access to a limited number of security researchers in the AI community.