Block Launches Dessa: Full-Duplex Voice Model Runs Locally on MacBook M4 Max

Executive Overview

On October 24, 2026, Block, the company led by Jack Dorsey, announced Dessa, a full-duplex voice model that runs locally on consumer hardware such as the MacBook M4 Max. Trained entirely on human speech and a single NVIDIA GPU, Dessa is currently available only to safety researchers via Hugging Face, with no commercial pricing announced. The model represents a architectural shift by removing the text pipeline from voice processing, embedding emotion directly into the model to produce natural laughter, pauses, and crying without external voice tags.

Deep-Dive Features & Architecture

Dessa is built on a full-duplex architecture that allows it to speak and listen at the same time, similar to OpenAI's GPT-Live model released earlier this year. The core innovation is the removal of the text pipeline that current models like ElevenLabs v4 and Fish Audio Drama 3 rely on, which use voice tags to drive emotion. In Dessa, emotion is embedded directly into the model itself, producing natural laughter, pauses, and crying without external tags.

The Block team says current models "convert your voice to text and discard a huge amount of information that text alone does not carry," and that the most powerful AI models "don't hear how you feel." This approach addresses the loss of tone, rhythm, and emotion in speech-to-text-to-speech systems. The model is trained entirely on human speech and a single NVIDIA GPU, making it lightweight enough to run on a MacBook M4 Max or better.

There is no public access to Dessa currently; Block has issued a call for safety researchers to test it, particularly in scenarios involving user deception through voice agents. This reflects early awareness of voice impersonation risks in financial applications, where Block owns payments business Square, which already uses voice AI to process orders.

Benchmark & Competitive Performance

No official benchmark figures have been published for Dessa yet, but qualitative comparison with competing models reveals fundamental differences: ElevenLabs v4 and Fish Audio Drama 3 rely on voice tags to drive emotion, while Dessa integrates emotion into the architecture itself. Dessa also runs locally on consumer hardware, whereas competing models rely on the cloud. This gives Block a low operating cost advantage for Square and financial services applications, without the need for cloud API calls.

Industry Impact & Enterprise Adoption

The absence of explicit Arabic support in Dessa means Arabic developers cannot currently rely on it for Arabic conversational applications. However, the local operation model on MacBook M4 Max opens the door to low-cost Arabic voice experiments if Block adds language support later. Token efficiency does not apply here because the model is purely voice-based and does not rely on token pricing, reducing development costs for startups in the Arab world that struggle with cloud API prices. Practical use cases include voice agents for e-commerce and local financial services, provided Arabic support becomes available. Until then, Dessa remains an architectural reference for building voice models.

Conclusion

Block's Dessa marks a significant step toward emotionally intelligent, locally run voice AI. By eliminating the text pipeline and embedding emotion directly into a full-duplex architecture, Dessa offers a glimpse of more natural human-machine voice interaction. While commercial availability and Arabic support remain unconfirmed, the model's efficiency and local operation on consumer hardware signal a promising direction for cost-effective voice applications in finance and beyond.

Media Source: TechCrunch AI | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis