Block Unveils Dessa Speech AI That Runs Locally on an M4 Max Mac
Block has revealed Dessa, a native speech-to-speech AI model trained entirely on human conversation that runs locally on high-end Apple hardware. Unlike conventional voice assistants that daisy-chain text transcription and synthesis, Dessa processes spoken audio directly on the machine without routing data to the cloud.
Native speech without cloud roundtrips
Dessa processes and generates audio directly rather than converting sound into text and back again.
Most digital voice assistants rely on a three-step pipeline: transcribing spoken input into text, running that text through an LLM, and feeding the written answer into a text-to-speech engine. The process introduces unnatural latency and strips away conversational subtleties like tone, hesitation, and cadence.
Block took a fundamentally different engineering path with Dessa. By handling raw audio directly end-to-end, the architecture preserves acoustic cues and cuts response delay dramatically. Because the model was trained exclusively on organic human dialogues, conversational pacing and interruptions sound far more natural than synthetic alternatives.
Block Dessa: requirements and technical profile
| Spec | Detail |
|---|---|
| Developer | Block |
| Architecture | Direct end-to-end speech-to-speech |
| Training dataset | Exclusively human conversation audio |
| Minimum hardware requirement | MacBook Pro with Apple M4 Max or higher |
| Execution mode | Fully local (offline) |
The steep hardware price of running voice AI locally
The standout technical achievement here is privacy: no cloud servers are required to converse with the system. That makes it an enticing prospect for users wary of corporate data retention, particularly given recent concerns about conversational logs where tools like ChatGPT locked chats lingering data highlighted the risks of centralized cloud storage.
That independence comes at a severe computing cost. Block specifies that running Dessa locally demands a MacBook Pro configured with at least an M4 Max chip. Heavy unified memory bandwidth and substantial RAM overhead are likely the gatekeepers keeping older or entry-level M-series machines out of the loop.
Running an end-to-end speech model locally is an enormous technical feat, but the M4 Max requirement is a reminder of how demanding real-time multimodal audio processing remains. Until optimization techniques or smaller quantized variants emerge, Dessa will remain a proof-of-concept playground for high-end workstation owners rather than a practical daily utility for the masses.
Availability and hardware requirements
Block has not yet announced broader public distribution timelines, open-source licensing terms, or whether weights will be freely downloadable. What is clear right now is the barrier to entry: only users with Apple's highest-tier silicon, specifically an M4 Max MacBook Pro or equivalent future workstation, meet the baseline requirements to run Dessa locally.
Frequently asked questions
What hardware do you need to run Block Dessa locally?
According to Block, you need at least an Apple MacBook Pro equipped with an M4 Max processor or better to handle the model's compute and memory requirements.
How is Dessa different from typical voice assistants?
Dessa operates as a true speech-to-speech model, taking in audio and outputting voice directly without first converting user input into intermediate written text.
Sources
- Block, Yerel Olarak Çalışabilen Konuşmadan Konuşmaya Yapay Zeka Modeli Dessa'yı Duyurdu. @Techmeme, 2026. https://x.com/Techmeme/status/2114001747572396276