Evidence at a glance
The Change Is More Than Better Turn-Taking
Block’s AI research team announced Dessa on October 23, 2026, as a research preview. Dessa is a full-duplex speech-to-speech model: it takes audio in, produces audio, and can keep listening while it speaks. Block says it was trained only on human conversation audio, without text or pretrained models. That makes it more than a more natural voice interface. It is also an architectural claim about what representation a speech model should learn from.
A common voice system chains speech recognition, a text-based conversational model, and speech synthesis, with text serving as the intermediate interface. Dessa claims to bypass that textual layer and generate audio directly from audio. The distinction for technical leaders is therefore not simply whether a model can listen and speak at once. Other full-duplex systems have shown that. The central claim is whether a model must first compress a conversation into text before it can understand and answer it.
What Bypassing Text Might Preserve—and Lose
A transcript can record what someone said without fully preserving how they said it. Laughter, sighs, sarcasm, and changes in volume may be weakened or disappear in text. Block argues that learning directly from conversational audio could make those cues more available to the model, helping it learn when to interrupt or offer short acknowledgments such as “yeah” or “got it.” Public demos show laughter, backchannels, and interruptions, the kinds of behavior this design is meant to address.
But evidence is still needed to connect those behaviors to the design. Block has not published Dessa’s architecture, audio encoding method, or training objective. Nor has it released controlled comparisons showing that speech-only training caused the demonstrated behavior rather than some other undisclosed factor. Skipping text may avoid losing acoustic information, but it also removes an intermediate representation that engineers can inspect. Logging, diagnosing errors, and reviewing conversation content could become harder. Preserving more sound should not be taken to mean understanding better.
Moshi Makes the Difference in Approach Concrete
Kyutai’s Moshi is a useful comparison because it also supports full-duplex conversation, but starts from a text-language-model backbone and then predicts time-aligned text and audio tokens. Full duplex, then, is not a capability unique to Dessa. The difference is what each system puts at the start of its speech-generation route. Moshi incorporates a text model, while Dessa presents the use of neither text nor pretrained models as its defining choice.
The disclosures are also uneven. Moshi’s model card lists 8 billion parameters, 7 million hours of pretraining audio, and latency of about 160 milliseconds in theory and 200 milliseconds in practice. Dessa has published neither a parameter count nor a specific latency figure, saying only that inference is faster than real time. These details show a difference in approach and in the evidence available, but they do not establish which model is better at conversation. Without public weights or benchmarks, outside teams cannot reproduce a comparison using the same tasks, devices, and measurement protocol.
A Large GPU Count Gap Does Not Prove Efficiency
Block says Dessa was trained on one NVIDIA GB200 NVL72 system, a rack-scale machine containing 72 Blackwell GPUs. Moshi’s disclosures cite 1,016 H100 GPUs. It is tempting to read these figures as an answer about training cost, but they refer to different GPU generations, and Block has not published Dessa’s training duration or total compute. GPU counts alone cannot show which model used less compute or trained more efficiently.
The inference claims also need to be separated into what is stated and what is measured. Block says Dessa can run locally on a MacBook Pro M4 Max or better and operates faster than real time. Public materials give no millisecond latency, memory requirement, or standardized test result, so this is a vendor-reported capability rather than a reproducible performance metric. In one demo, a user asks the model to speak louder and it raises its volume. That shows a specific response in a clip, not stability across devices, environments, or conversation tasks.
A Research Preview Is Not Yet a Deployment Basis
Block also says Dessa learned language with hundreds of times fewer words than large language models require, but has not disclosed the total amount of training data, its method for counting words, or the comparison baseline. That data-efficiency claim cannot currently be checked independently. The source, licensing, and language coverage of the training audio have also not been made public. For teams evaluating the model, these are not peripheral details. They affect whether results can be reproduced, which users the model may serve, and whether it can fit into a real product workflow.
Dessa has no public weights, technical paper, benchmark scores, or general-purpose API. It remains a research preview, with access through a waitlist and limited early applications. Technical teams can treat it as a research hypothesis and test whether speech-only training improves interruption timing, backchannels, and vocal nuance. A practical decision should wait for fuller model and evaluation disclosures. Until then, claims of local execution, natural demos, and a smaller GPU count do not justify replacing an existing voice pipeline or making a purchasing decision.