Block Debuts Local Speech-to-Speech AI Model Named Dessa
The new model runs locally on high-end hardware and is currently being distributed to a small group of AI safety researchers.
Updated on Oct. 23, 2026 in Artificial Intelligence
Live Poll
Do you trust artificial intelligence models that mimic human social behavior and awkwardness?
Block has announced the release of Dessa, a new speech-to-speech artificial intelligence model trained on human conversation. This model is designed to run locally on hardware as powerful as a MacBook Pro M4 Max.
Why it matters
The model represents a shift toward edge-based AI, allowing conversational processing to occur on local hardware rather than in the cloud. Block cited curiosity, safety, and human-AI understanding as the primary drivers for this research-stage development.
Dessa utilizes speech-to-speech processing and is trained from scratch on human speech recordings. It is built to operate locally, requiring at minimum a MacBook Pro M4 Max or equivalent hardware for execution.
The players
Block
A financial technology company that operates platforms like Square and Cash App and is now expanding into local AI research.
The details
Dessa is a speech-to-speech model, meaning it bypasses traditional text-based intermediate steps to process auditory input directly into conversational output. By training the system from scratch on human speech recordings, the model aims to capture natural cadence and interaction patterns. It is optimized to perform these tasks locally, eliminating the need for remote server communication, provided the user has hardware capable of handling the computational load of an M4 Max chip.
Timeline
October 23, 2026: Block announced the development and capabilities of the Dessa AI model.
The Tech Race
Dessa marks a departure from large-scale, cloud-dependent speech models, positioning itself within the growing movement to bring complex generative AI to the edge. It competes in a research space where localized processing is increasingly prioritized for latency and privacy reasons.
Developers and AI safety researchers can expect limited early access to the model, though general public availability has not been established. Users will need hardware with the computational overhead of an M4 Max chip to run the model locally.
The takeaway
The development of Dessa highlights a growing focus on local-first AI, though the project remains in an early research phase. Watch for further findings from the safety researchers currently evaluating the model to see if its speech-processing approach offers tangible safety advantages over current cloud models.
Further reading
For broader trends in model development, visit Artificial Intelligence.
Live Poll
Do you trust artificial intelligence models that mimic human social behavior and awkwardness?
