Block unveiled Dessa on October 23, 2026, introducing a speech-to-speech AI model that it says can listen while speaking and run locally on a MacBook Pro M4 Max or better. The model comes from a small research lab inside the payments company, with a focus on conversational timing, expressive voices and learning directly from human speech. TechCrunch reported the announcement the following day.
Dessa is not generally available. Block’s announcement page offers recordings, a waitlist and an invitation for a small group of AI safety researchers to get early access. It does not provide a public model download or an API that developers can start building against.
That distinction matters. Block has announced an interesting research system, but the available evidence cannot yet tell developers whether its conversational performance, hardware requirements or safety behavior will hold up outside company-selected demonstrations.
What Block Has Actually Released
Block’s public release consists primarily of an explanation of Dessa and recordings intended to demonstrate its capabilities. The company says those recordings capture live conversations, that nobody steers the model, and that subtitles were added afterward. Those are Block’s descriptions of its demonstrations, not independently verified test conditions. The primary page also identifies the team as a Block research lab founded by former members of the Toronto deep learning startup Dessa.
The access situation is more limited than the word “release” can suggest:
| Item | Publicly documented status |
|---|---|
| Conversation recordings | Available on Block’s Dessa page |
| General access | Not available in the supplied reporting |
| Waitlist | Open on the announcement page |
| Safety testing | Early access offered to a small group of researchers |
| Downloadable weights or public API | Not provided on the announcement page |
| Model size, license and public release timetable | Not specified in the opened primary material |
TechCrunch also reported that Dessa was not publicly accessible at the time of its coverage.
I’d treat this as a research unveiling, not an open-source launch or a deployable product announcement. Local execution describes where a model can run. It says nothing by itself about whether its weights are available, what license applies or who has permission to use it.
Full-Duplex Speech Changes the Conversation Loop
Dessa’s central technical claim is full-duplex operation: audio can enter and leave the system at the same time. Block says the model can keep listening while it talks, choose when to interrupt and produce brief acknowledgments such as “yeah” or “got it.” Its description presents these behaviors as things learned from human conversation.
That differs from a strictly turn-based voice interface, where the system waits for the user to finish before producing a response. Full-duplex operation creates room for more fluid exchanges:
- Mid-sentence feedback: A short acknowledgment can signal that the system is following.
- Overlapping speech: A user can begin speaking before the assistant finishes.
- Changing direction: New speech can potentially affect an exchange already in progress.
The practical challenge is deciding what an overlap means. A brief acknowledgment should not necessarily stop the assistant. A correction or a request to stop usually should. Simultaneous input and output make those interactions possible, but reliable interruption handling still needs separate evaluation.
Block also describes a different training approach. It says Dessa was trained from scratch on human conversation, with “No text, no pretrained models.” The model takes audio in and produces audio out, with knowledge learned from speech. Block’s account contrasts this with systems that turn speech into text before generating a spoken response.
The distinction is consequential because a transcript alone leaves out information carried by pitch, rhythm, volume, pauses and emphasis. The same words can communicate reassurance, irritation or sarcasm depending on delivery.
Speech-only training does not, however, establish how well Dessa reasons, retrieves facts or follows complicated instructions. Block itself raises open research questions about learning facts, acquiring new languages and in-context learning. Its announcement does not supply evaluations answering them.
For readers following LLM releases, Dessa therefore belongs in a different comparison category. A speech-trained conversational model needs evidence about language competence and conversational behavior, not simply a convincing voice sample.
Local Inference Is a Hardware Claim, Not a Deployment Plan
Block says Dessa runs locally on a MacBook Pro M4 Max or better, with inference faster than real time. It separately says training used one NVIDIA GB200 NVL72 system. Both statements come from Block’s announcement, and neither is accompanied there by a reproducible performance report.
Training and inference are different workloads. The NVIDIA system describes the equipment Block says it used to create the model. The MacBook claim concerns running the trained model. Neither statement supplies the missing details developers would need to size a deployment.
The M4 Max reference is useful, but incomplete. Block does not specify required memory, the exact laptop configuration, sustained performance, power consumption or how much capacity remains for other applications. It also does not define which machines count as “better.”
That makes “runs on consumer hardware” more specific than a vague on-device promise, but much less specific than a published minimum system requirement.
Faster Than Real Time Is Not a Latency Measurement
Faster-than-real-time inference generally means a system can process or generate audio faster than that audio takes to play. That is important for keeping a conversation moving, but it does not tell us how long a person waits for the first response.
End-to-end responsiveness also depends on audio buffering, microphone and speaker handling, processing chunks and decisions about when to speak. A model can have adequate processing speed while still pausing noticeably before answering.
Block provides no numerical conversational latency measurement on the announcement page. Its speed claim therefore cannot be converted into a response time in milliseconds.
My view is that this is the most important missing deployment detail after access. For a voice interface, a convincing sentence arriving late can be less useful than a simpler response arriving at the right moment. A useful evaluation would measure first-response delay, interruption response and performance during overlapping speech on a specified machine.
Local inference could reduce reliance on a network connection for the model itself and change operating costs. It could also support keeping conversational audio on a device. Those are potential benefits, not documented Dessa product guarantees: local execution alone does not establish a complete application’s privacy policy, telemetry behavior or total cost.
Expressive Speech Is Different From Proven Understanding
Block’s recordings are presented as examples of laughter, emotional delivery, changes in volume, sentence completion and conversational knowledge. One scenario involves a forgotten movie title; another features a surprised, emotionally expressive response. The company describes Dessa as naturally hearing and speaking emotion. Those claims and examples appear on its product page.
Three separate questions sit behind that presentation:
- Can the model produce a convincing emotional voice?
- Can it recognize emotion accurately from someone else’s speech?
- Can it respond appropriately when that interpretation matters?
Frequently Asked Questions
3 questions
1Can Anyone Download Dessa?
No public download is provided in the opened announcement. Block offers a waitlist and limited early access for safety researchers.
