Block Dessa: Can You Download the Voice Model Yet?
Block's Dessa voice AI promises speech-only, full-duplex conversation and local Mac inference. What is actually available, what remains unverified, and how it compares with Moshi.
Block introduced Dessa on October 23, 2026 as a research preview of a voice AI designed for conversations that do not feel like a sequence of dictated commands. The headline is not a public app launch: Block says the model learned to converse from recorded human speech alone, without text data or pretrained models. Its official page includes conversation recordings, but does not provide a usable public model download, API, price, or release timetable.
What Dessa does differently
Dessa is described as speech-to-speech and full-duplex: it can keep listening while talking, interrupt, acknowledge a speaker mid-sentence, and respond to cues such as laughter, sarcasm, pauses and changes in tone. Block says it trained the model from scratch using human conversations, instead of routing speech through a text-first pipeline. These are the company's architectural and behavioral claims; the released samples are selected demonstrations, not independently repeatable tests.
The company says its training used one NVIDIA GB200 NVL72 rack. NVIDIA specifies that the system contains 72 Blackwell GPUs and 36 Grace CPUs. That clarifies the scale of the hardware Block named; it does not independently establish Block's training cost, duration, dataset size or efficiency. Block has not released the technical detail needed to reproduce those results.
Can developers actually use Dessa yet?
Dessa availability: what is verified
| Question | Status as of October 26, 2026 |
|---|---|
| Official conversation recordings | Available on Block's Dessa page; company-selected examples |
| Public model download or weights | Not provided; page says launch links are coming soon |
| API, SDK, or API price | No public documentation or rate card on the launch page |
| License or commercial-use rights | Not published on the launch page |
| Early access | Block invites a small group of AI safety researchers to apply |
| Local Mac support | Block claims MacBook Pro M4 Max or better; no public setup to verify |
A model being described as locally runnable does not mean anyone can install it now, nor does it establish offline privacy guarantees. Until weights, setup steps, licensing and data-handling documentation are public, treat Dessa as a research preview rather than a deployable voice-agent component.
Dessa versus an available full-duplex alternative
Architecture and access, not a performance ranking
| Dimension | Block Dessa | Kyutai Moshi |
|---|---|---|
| Input and interaction | Speech-to-speech; full-duplex (Block description) | Speech-text full-duplex dialogue (Kyutai documentation) |
| Role of text | Block says no text or pretrained models in training | Predicts text tokens alongside audio |
| Public model access | No downloadable weights or API shown | Public code, models and installation instructions |
| Mac deployment | M4 Max or better claimed; setup unpublished | Official MLX inference implementation documented |
For a team shipping a voice agent this month, model availability matters more than the most impressive demo clip. Moshi offers a reproducible starting point for prototyping on a Mac, while Dessa is worth tracking as a different research approach to learning speech and conversational timing. That is not a claim that Moshi is more accurate or that Dessa is faster: the companies have not published directly comparable, independently run evaluations.
What to check before building on Dessa
- Watch the official samples for interruption handling and timing, but note that the clips are selected by the developer.
- Wait for model weights or an API and a written license before planning a commercial integration.
- Require tested memory use, latency percentiles, supported languages, and repeatable benchmarks on named hardware.
- Test corrections, overlapping speakers, background noise and emotionally sensitive conversations with human review.
- Check recording consent, retention and security documentation before exposing customer audio to any new system.
Block says it is asking outside AI safety researchers to identify risks the team may have missed. That matters for voice interfaces because a convincing emotional delivery can make an incorrect answer sound more trustworthy. No public safety evaluation, general release date or integration with Block's payment products is established by the official page.
Questions people ask
Can I download Dessa or use it in an app today?
No public Dessa model download or developer API is documented on Block's launch page as of October 26, 2026. The page offers recorded demonstrations and says launch links are coming soon. Block is inviting a small group of AI safety researchers to request early access, but that is not general availability. Do not treat audio sample downloads as model weights or an integration SDK.
Does Dessa really run on a MacBook Pro with an M4 Max?
Block says Dessa can run locally on a MacBook Pro M4 Max or better, but it has not published a downloadable build, installation guide, required memory, or reproducible performance measurements. That is a company hardware claim, not a verified setup that readers can reproduce today. Wait for official model files and supported-device requirements before buying hardware specifically for Dessa.
Is Dessa open source, and what does its API cost?
Neither public model weights, a distribution license, nor API documentation or pricing appears on Block's Dessa research-preview page as of October 26, 2026. The company describes its training approach and shares demonstrations, but does not establish a commercially usable release. An announcement that a model can run locally is not evidence that its weights are available or licensed for commercial deployment.
How is Dessa different from Kyutai's Moshi voice model?
Both are described as full-duplex systems that can listen while speaking. Block's distinctive claim is training Dessa from scratch on human speech without text or pretrained models; independent reproduction is not yet possible. Kyutai's official Moshi repository describes a speech-text architecture that predicts text alongside audio and provides downloadable models and an MLX implementation for Macs. There is no comparable public benchmark proving Dessa performs better.
Dessa is a noteworthy speech-only, full-duplex research preview, not a publicly deployable voice model. The strongest current story is the gap between Block's ambitious training and local-inference claims and the missing weights, API, benchmarks and licensing. Builders should evaluate reproducible alternatives today and revisit Dessa when the company publishes verifiable access and technical documentation.
Explore related Make Better resources
- OpenAI Prompt-to-Voice: API, Cost and Limits— Related but distinct shipping voice API and pricing intent
Sources & useful resources
- Block: Dessa official research preview and demonstrations— Primary company statement: speech-only training, full-duplex design, Mac M4 Max claim, safety-researcher access; page has no public release date
- Kyutai: Moshi official open-source repository— Primary model documentation for an already available full-duplex speech-text alternative and MLX inference
- NVIDIA: GB200 NVL72 specifications— Official rack specification, 72 Blackwell GPUs; verifies what the named training hardware contains, not Block's actual use
- TechCrunch: Block's Dessa announcement, October 24— Contemporaneous reporting of Block's October 23 announcement; secondary dating corroboration only