Block Unveils Dessa, Speech-Only Model

Dessa is a provocative architecture bet, but developers get an impressive demo—not yet a usable platform, with no public weights, benchmarks, pricing, or access.

• –Native speech-to-speech enables interruption, backchanneling, emotion, and more natural turn-taking than cascaded voice stacks.
• –Speech-only training challenges the assumption that text is essential for learning language and facts.
• –Local M4 inference hints at private, edge-friendly deployment, though Block provides few performance details.
• –Inviting safety researchers is sensible because expressive, autonomous voice behavior creates unfamiliar safety risks.
• –The lack of model size, dataset details, evaluation results, and a public release makes comparison with GPT Realtime or Gemini Live difficult.