Block releases speech-to-speech AI model Dessa for local use

Builders evaluating local voice agents can test Dessa’s speech-to-speech approach on M4 Max hardware to measure latency and emotional fidelity against text-first pipelines. A conditional scenario: a developer prototyping an on-device companion app would first confirm the Mac meets the stated hardware floor, then run the provided conversation examples to check whether the model maintains timing without cloud round-trips. The evidence does not establish licensing terms, weight availability, or performance on other silicon.

Operators considering voice agents for sensitive user interactions should review the safety-researcher access program and any data-handling commitments before internal pilots. A conditional scenario: a product team evaluating on-device voice for health-adjacent features would first verify whether their organization qualifies for the small-group access and confirm that all audio stays on controlled hardware. The supplied evidence does not establish retention policies or commercial licensing.