Dessa
A conversational AI model launched by Block, trained entirely from scratch on human conversation data. It focuses on social pragmatics and natural conversational experience rather than general-purpose knowledge question answering.
Timeline (last 90 days)
Block's AI lab released a full-duplex speech-to-speech model named Dessa.
Dessa can run locally on a MacBook Pro M4 Max.
Dessa's training data comes solely from human conversation audio, with no text transcripts or annotation corpora.
An AI voice demo of Dessa on Reddit sparked heated discussion, with some users marveling at its realism and some questioning whether it was a scam.
Dessa's voice demo includes dialogue simulating a home scenario, such as "Wolfie's fine, honey. Wolfie's just fine. Where are you?" as well as "he's smart" and "Work is tough," with natural and coherent intonation.
Some users speculate that Dessa's technical lineage may be moshi -> sesame -> dessa, but this claim has not been officially confirmed.
Dessa belongs to block.xyz, not an OpenAI project.
Dessa has been specifically optimized for speech naturalness, response latency and conversational fluency.
The Block team released Dessa, which it officially calls "the most realistic full-duplex speech model to date".
Dessa's release received 288 upvotes and 113 comments on Hacker News.
8 more timeline events
All Facts (18)
Dessa can run locally on a MacBook Pro M4 Max.
50%UnverifiedDessa's training data comes solely from human conversation audio, with no text transcripts or annotation corpora.
50%UnverifiedDessa belongs to block.xyz, not an OpenAI project.
50%UnverifiedDessa's voice demo includes dialogue simulating a home scenario, such as "Wolfie's fine, honey. Wolfie's just fine. Where are you?" as well as "he's smart" and "Work is tough," with natural and coherent intonation.
50%UnverifiedSome users speculate that Dessa's technical lineage may be moshi -> sesame -> dessa, but this claim has not been officially confirmed.
50%UnverifiedDessa has been specifically optimized for speech naturalness, response latency and conversational fluency.
50%UnverifiedThe Dessa model was trained entirely from scratch on human conversation data, without relying on web text, books, or code.
50%UnverifiedDessa completely abandons non-conversational corpora, which naturally limits its breadth of knowledge but may give it an advantage in conversational naturalness.
50%UnverifiedDessa's pure conversation training approach aims to make the model focus more on social pragmatics, such as turn-taking, emotional responses, and continuation of everyday topics.
50%UnverifiedBlock did not disclose the number of model parameters or the scale of training data for Dessa.
50%UnverifiedDessa's lead is Ragavan Thurairatnam.
50%UnverifiedDessa's goal is not to be an all-knowing assistant but to simulate a natural, even slightly awkward, social conversation experience.
50%UnverifiedDessa's conversational style is officially described as "like chatting with a new friend," with exploration and uncertainty, and its replies sometimes seem overly literal or lacking contextual connection.
50%UnverifiedBlock's AI lab released a full-duplex speech-to-speech model named Dessa.
50%UnverifiedAn AI voice demo of Dessa on Reddit sparked heated discussion, with some users marveling at its realism and some questioning whether it was a scam.
50%UnverifiedThe Block team released Dessa, which it officially calls "the most realistic full-duplex speech model to date".
50%UnverifiedDessa's release received 288 upvotes and 113 comments on Hacker News.
50%UnverifiedBlock (led by Jack Dorsey) released an AI model named Dessa.
50%