Dessa review (2026): an honest first look at Block's full-duplex speech model — what it does, how it was trained, availability, and who it fits.
Dessa is one of the most ambitious conversational speech models announced to date: full-duplex, trained only on human speech, emotionally expressive, and able to run locally on a high-end Mac. It is also not generally available — waitlist-only, with no independent benchmarks, pricing, or release date — so this is a first look, not a usage verdict. The score blends high marks for the technology with low marks for availability and proof; and if your real goal is producing and publishing content, a conversation model isn't that tool.
Dessa is a full-duplex speech-to-speech model unveiled in October 2026 by Block's small in-house AI research lab — the payments company co-founded by Jack Dorsey, whose lab reports to him and was built partly from the team behind Dessa, the Toronto deep-learning startup Square (now Block) acquired in 2020. The model is named after it. This is an honest first look, not a hands-on verdict: at announcement Dessa is waitlist-only, with early access limited to AI safety researchers, so the scores below weigh a genuinely novel piece of technology against the fact that almost nobody can use it yet.
I run Kompozy, a content generation and publishing engine, so the disclosure is upfront — and it shapes the one thing I want to be clear about before any scores: Kompozy and Dessa are not competitors. Dessa is a model you converse with. Kompozy produces and publishes content. I'm reviewing Dessa on its own terms as a speech model, and where it's impressive I say so. The standout is the training recipe: Block says Dessa learned entirely from human conversation, with no text and no pretrained models, on a single NVIDIA GB200 NVL72, and that it carries emotion, sarcasm, laughter, and timing natively rather than flattening speech into text first. Full-duplex — listening and talking at the same time, with learned interruptions and backchanneling — is the hardest part of natural dialogue, and it's the explicit design goal here.
The caveats are about availability and proof, not ambition. There are no independent benchmarks; every quality, speed, and efficiency figure is vendor-reported, and the product is days old. There's no pricing, no stated license, and the flagship demo pairs Dessa with a third-party model, GPT-Live-1, that Block says is unaffiliated and hasn't endorsed it. Everything below reflects Dessa as of 2026-10-23, drawn from Block's own announcement and demo page; treat the superlatives ("most authentic," "world's first") as marketing until third-party tests land, and verify any detail on block.xyz/dessa before you build around it.
Dessa is a full-duplex, speech-to-speech conversation model: audio in, audio out. Because it is full-duplex, it listens and speaks at the same time and learned from real conversation when to cut in and when to backchannel with short acknowledgments like "yeah" or "got it." Block says it was trained from scratch on human conversation alone — no text, no pretrained models — on a single NVIDIA GB200 NVL72, and that it natively handles emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness, with inference that runs faster than real time so replies arrive without an awkward pause. It reportedly runs locally on a MacBook Pro M4 Max or better, and lead researcher Ragavan Thurairatnam said it can also run on consumer GPUs. It is a model, not an application. Dessa holds a spoken conversation; it does not write copy, generate images or video, caption anything, assemble a carousel or a blog, or publish to any platform. At announcement it is not generally available: Block opened a public waitlist and offered early access to a small group of AI safety researchers through an application requiring an institutional email, saying launch links are "coming soon." There is no published pricing, no stated open-source license, and no general release date. The public demos are audio recordings, including a conversation with a third-party model (GPT-Live-1) that Block states is not affiliated with it.
Dessa, in the state it's announced in, is for developers and researchers building two-way voice experiences — a conversational agent, a companion, a live coach, a voice interface — who want expressive, low-latency, on-device speech and are willing to wait for access or apply through the researcher program. If you care about the frontier of speech-native model training, a from-scratch, speech-only model with researcher early access is directly interesting. For a creator whose job is producing posts, it is a weaker fit twice over. First, it isn't available: you can't reliably plan content around a model behind a waitlist. Second, and more fundamentally, it's the wrong category — Dessa generates spoken conversation, not the captioned clips, carousels, quote graphics, blogs, newsletters, and scheduled posts that make up a content operation. It's worth watching, and worth joining the waitlist if voice products are your lane, but it is not a tool for shipping on-brand content across platforms.
| Dimension | Score | Why |
|---|---|---|
| Naturalness & expressiveness (claimed) | 4.3 / 5 | Demos show emotion, sarcasm, laughter, sighs, and a wide intonation range; audio-native rather than text-first. Vendor demos, not independent tests. |
| Full-duplex conversation | 4.2 / 5 | Listens and talks at once with learned backchanneling and interruptions — the core design goal and the hardest part of natural dialogue. |
| Latency / real-time performance | 4.0 / 5 | Faster-than-real-time inference is claimed, which is what keeps live conversation from feeling awkward (vendor-reported). |
| On-device / local operation | 4.0 / 5 | Runs locally on a MacBook Pro M4 Max or better and reportedly on consumer GPUs — strong for privacy and latency, though it needs high-end hardware. |
| Training approach / efficiency | 3.8 / 5 | Trained from scratch on speech alone, no text or pretrained models, on a single GB200 NVL72 from far fewer words than LLMs — novel if it holds up. |
| Documentation / transparency | 2.5 / 5 | A demo page and a set of claims, but no model card, technical paper, or published evaluation at announcement. |
| Availability / access | 1.8 / 5 | Waitlist-only with early access limited to AI safety researchers; launch links are "coming soon." |
| Pricing clarity | 1.5 / 5 | No pricing, commercial terms, or stated license announced. |
| Independent benchmarks / track record | 1.5 / 5 | No third-party evaluation; all figures are vendor-reported and the product is days old. |
There is nothing to reconcile here yet: Block announced no pricing, no license, and no commercial terms for Dessa. That alone caps the "value" question — you can't assess a price that doesn't exist, and a waitlist product with no stated terms is, for now, something to watch rather than budget for.
What the architecture implies is worth noting. Because Dessa is pitched as running locally on a MacBook Pro M4 Max or better (and reportedly on consumer GPUs), the real cost of an on-device model tends to be hardware and engineering rather than a per-use bill — if it ships that way, you'd own the compute and avoid metered cloud charges, but you'd also need capable hardware and the work to build a product around it. None of that is confirmed as a commercial offering.
Scored as value, Dessa is simply incomplete: promising technology with no price, no license, and no general availability. Revisit this once Block publishes terms. If the question underneath "is it worth it?" is really "will this help me produce and publish more content?", the honest answer is that a conversation model isn't where that spend belongs — that's a content-engine cost, a different line item entirely.
| Use case | Fit | Why |
|---|---|---|
| Building a live voice agent or conversational companion | Strong | Full-duplex, low-latency, expressive speech is exactly the primitive a two-way voice product is built on. |
| Researching speech-native model training | Strong | A from-scratch, speech-only model with a researcher early-access program is of direct interest to the field. |
| Private, on-device voice experiences | OK | Local operation on a Mac means no cloud dependency, useful for privacy-sensitive apps — once access opens up. |
| Prototyping voice UX you can rely on today | Weak | It is waitlist-only; you can't reliably build on a model you can't access yet. |
| Batch narration or voiceover files at volume | Weak | Dessa is a two-way conversation model, not a bulk text-to-speech pipeline — and it isn't available. |
| Generating social posts, clips, or video from an idea | Weak | It writes no copy, makes no visuals, and produces no content; that's a different category of tool. |
| Scheduling and publishing content across platforms | Weak | No distribution layer at all — it's a speech model, not a publishing engine. |
Calling Kompozy a Dessa competitor would be dishonest — they govern opposite ends of the job. Dessa governs a live spoken exchange: tone, timing, when to cut in, when to say "yeah." Kompozy governs a brand across produced, published content: a Persona Brief holds voice, tone, and banned words consistent not on one conversation but across captions, carousels, quote graphics, short-form clips, blogs, and newsletters, and HyperFrames renders every layout pixel-exact. One is about sounding human in real time; the other is about showing up on-brand, on schedule, everywhere.
Where the two actually meet is upstream of Kompozy: conversation is a content source. A recorded podcast, interview, or voice session — whether voiced by a model like Dessa down the line or by you and a guest today — is raw material Kompozy turns into finished output. It cuts the audio into captioned Clipped Shorts, pulls the sharpest lines into Quote Graphics, drafts a blog and a newsletter from the transcript, and re-presents a key answer as a face-locked Persona Short, then schedules and publishes the set across the eight social platforms plus blog and email. If you're weighing Dessa because you want to "make more content," that's the honest redirect: the model makes the talk sound real; Kompozy turns talk into a published presence.
Not generally, as of this review. At announcement Block opened a public waitlist and offered early access to a small group of AI safety researchers through an application requiring an institutional email, saying launch links are coming soon. There is no general release date, pricing, or stated license yet.
That is Block's claim: it says Dessa was trained from scratch on human conversation alone — "sounds, no text, no pretrained models" — on a single NVIDIA GB200 NVL72, learning language from far fewer words than large language models. It's a genuinely novel recipe, but there are no independent benchmarks yet, so treat the figures as vendor-reported.
Block says Dessa runs locally on a MacBook Pro M4 Max or better, and lead researcher Ragavan Thurairatnam said it can also run on consumer GPUs. That points at an on-device model, though full deployment details weren't published at announcement.
Both are described as full-duplex voice models, but Dessa's pitch is that it was trained only on human speech (no text) and runs locally on a Mac. Block's demo includes a conversation between the two, while noting GPT-Live-1's developer is unaffiliated and hasn't endorsed Dessa. Without shared independent benchmarks, a head-to-head quality verdict isn't possible yet.
Block did not announce an open-source license, pricing, or commercial terms at launch. Confirm the current status on block.xyz/dessa before assuming either.
No. Dessa is a conversation model — it speaks and listens. It does not write posts, generate images or video, caption clips, assemble carousels or blogs, or publish anywhere. A content engine like Kompozy covers that half: it turns a recorded conversation into captioned clips, quote graphics, a blog, and a newsletter, then schedules and publishes across the eight social platforms plus blog and email.
Dessa was built by Block's in-house AI research lab, which reports to Jack Dorsey and was founded partly by former members of Dessa, the Toronto deep-learning startup Square (now Block) acquired in 2020. Lead researcher Ragavan Thurairatnam leads the effort; the model is named after the original startup.