← Back to Resources
Voice AIOctober 26, 2026

Block Dessa: Full-Duplex Voice AI That Fits on Your Desk

Block—the company behind Square and Cash App—stepped into voice AI in late October 2026 with Dessa, a research preview that, according to TechCrunch and posts on X, is a speech-only full-duplex voice model capable of running locally on an M4 Max-class Mac. No transcription pipeline, no cloud round-trip, no turn-taking stumbles: just speech in, speech out, on hardware you own.

What shipped: speech-native, local-first

As reported on October 24, 2026, the Dessa research preview makes two claims that matter together:

  • Speech-only, full-duplex. The model consumes and produces raw speech—listening while talking, handling interruptions naturally—rather than chaining transcription → text LLM → synthesis.
  • Local execution. Demonstrated running on an M4 Max-class Mac, which puts conversational voice latency in the "physics of your room" range and keeps audio private by default.

We previously covered the speech-to-speech approach from the model side in Block Dessa speech-to-speech—this page focuses on the local-first angle: what it means that a frontier-adjacent voice model now fits on a laptop.

Why the text hop is the problem

Almost every voice agent you have talked to is three models in a trench coat: a transcriber turns your speech into text, an LLM thinks about the text, a synthesizer turns the answer back into speech. Each hop adds latency, each transcription is a chance to lose tone, hesitation, or sarcasm, and the whole chain ships your audio to someone else's datacenter.

Speech-to-speech models collapse that stack. The reported trade-offs are real, though: speech-native models are harder to inspect (no text trace of what the model "thought"), harder to bolt tools onto, and the research-preview quality bar is unknown until independently tested. The cloud stack keeps major advantages in tool use—see ElevenLabs realtime voice agents for the state of that art, and Call4Me for phone-agent tooling built on it.

Why Block, and why local matters

Block's interest is not mysterious: payments and point-of-sale are voice-adjacent domains (busy counters, accessibility needs, merchant hardware already on-prem), and a company with Square's hardware footprint views local-first voice as an edge component, not a demo. For everyone else, the signal is architectural: if full-duplex speech models run on consumer silicon today, the "voice agents must live in the cloud" assumption has an expiry date.

Local-first voice changes three calculus points at once: privacy (audio never leaves the device), latency (no network in the loop), and cost (no per-second inference bills). The open question is capability—research previews from even well-funded labs start behind the production-hardened cloud stacks. For the local-model serving pattern on your own machine, see Llama for Windows, which brings the same local-first posture to text models.

How to track Dessa from here

  • Watch Block's official Dessa page (block.xyz/dessa) for the move from research preview to anything usable—an API, weights, or product integration.
  • Look for independent latency and interruption-handling tests; full-duplex claims live or die on the awkward overlap cases.
  • Compare against the speech-to-speech coverage on our earlier Dessa page and the cloud alternative in ElevenLabs realtime voice agent.
  • If you build voice agents today, design your stack so the speech model is swappable—the field is moving too fast to marry one architecture.

Block Dessa FAQ

What is Block Dessa?

Dessa is Block's speech AI research effort, unveiled as a research preview in late October 2026. As reported by TechCrunch and in posts on X, the first Dessa model is a speech-only, full-duplex voice model that can run locally on an Apple M4 Max-class machine—no text intermediary, no cloud round-trip required.

What does full-duplex mean for a voice model?

Full-duplex means the model speaks and listens at the same time, like a phone call—interruptions, backchannel acknowledgments, and overlapping speech all work without turn-taking arbitration. Most voice assistants are half-duplex: they stop hearing while they speak.

Does Block Dessa run locally?

As reported, the research preview demonstrated the model running locally on an M4 Max-class Mac. That is the headline claim: full-duplex conversational speech quality without a datacenter. Verify current hardware requirements and availability on Block's official Dessa page.

Block Dessa vs cloud voice agents from ElevenLabs or OpenAI?

Cloud voice stacks (see our ElevenLabs realtime voice agents coverage) transcribe speech to text, reason on text, then synthesize speech—latency and privacy costs included. Dessa's speech-native design skips the text hop entirely, trading the broader tooling ecosystem of text-based stacks for immediacy and local privacy.

Can I use Dessa in production?

Not yet, based on what has been announced. Dessa is described as a research preview from Block's research team—there is no general-purpose product API. Watch Block's official channels for availability changes.

AIXList is an independent AI tools directory. This article summarizes the Block Dessa research preview as reported by TechCrunch and posts on X in late October 2026; we are not affiliated with Block. Research preview capabilities and hardware claims should be verified against Block's official channels at block.xyz/dessa.