Comparison

Dessa AI vs Moshi: Speech-Only vs Open-Weights Full-Duplex

2026-10-28 · 7 min read · By DessaAI Daily

Dessa and Moshi are the two names that keep appearing in the same sentence for a reason: both are full-duplex, speech-to-speech systems built to talk and listen at the same time. But they come from opposite bets about how a talking machine should learn. Moshi starts from a powerful text language model and learns to speak. Dessa, Block says, starts from speech and never touches text at all.

What Moshi is

Moshi was created by Kyutai, the Paris-based research lab, and presented in 2024 as a realtime, full-duplex spoken dialogue system. Kyutai released Moshi's weights openly, which let researchers and developers download, inspect and build on it — a major difference from Dessa today. Moshi is built on Kyutai's Helium text language model, then extended into audio. In other words, text is in its foundation, even though the interaction is spoken.

That foundation gives Moshi a practical advantage: a text model's knowledge and language structure are already there before speech training begins. It also means Moshi inherits the strengths and limits of that approach — strong language modeling, with emotion and timing learned as an audio layer on top of a system that fundamentally knows language as tokens.

What Dessa does differently

Block's claim for Dessa, announced October 23, 2026, is more radical: trained from scratch on human conversation audio only, with no text and no pretrained models. As we covered in what Dessa AI is, Block says Dessa learned language from hundreds of times fewer words than text large language models need. If true, that is not just a different recipe — it suggests speech carries far more learnable signal per word than the field assumed, because every spoken word arrives with tone, timing, emotion and speaker intent attached.

Dessa was trained on a single NVIDIA GB200 NVL72 system and, Block says, runs faster than real time locally on a MacBook Pro with an M4 Max chip or better. In demos it laughs mid-sentence, backchannels, finishes a speaker's sentence, and holds a conversation with OpenAI's GPT-Live-1. Those demos are Block-selected recordings; no independent evaluation has been published yet.

Open vs closed: the biggest practical difference

Today, the comparison is not close on access. Moshi's open weights mean a developer can actually run it, test it, fine-tune relatives of it — NVIDIA's PersonaPlex, for example, is described as a Moshi fine-tune — and publish results others can check. Dessa has released no weights, no API, no parameter count and no benchmarks. There is no Dessa download. Early access goes first to a small group of AI safety researchers, and everyone else joins the waitlist described in our waitlist guide.

Openness cuts both ways, and Block's caution is part of the story. A voice model that sounds convincingly human raises impersonation and scam risks, which is why Block is putting safety researchers first. Kyutai chose open science; Block, a payments company whose products move money, chose a gated preview. Neither choice proves which model is better — but they tell you a lot about what each organization fears and values.

Emotion: learned from life, or learned from labels?

Both systems aim beyond flat text-to-speech. Moshi was designed for expressive, realtime spoken dialogue, with its own voice and conversational flow. Dessa's pitch is that emotion is not a feature added to speech — in its training data, emotion is speech. It learned sarcasm, laughter, sighs and frustration the way it learned words: by hearing them in real conversations, in context, thousands of times.

Without a head-to-head test using the same prompts, microphones and judges — which does not exist yet — any claim that one “sounds more human” is listening preference dressed as measurement. What can be said factually is narrower: their training philosophies differ, Moshi is inspectable today, and Dessa's speech-only claim is the bolder scientific statement that independent researchers have not yet been able to verify.

Which one matters for you?

If you are a developer or researcher who wants to build now, Moshi is the only one of the two you can actually obtain and run. If you are watching for a breakthrough in how machines learn language, Dessa is the experiment to follow: a model that may show how much of language lives in sound alone. And if you are a business user, neither replaces a supported product yet — Dessa is a preview, and Moshi is a research system, not a packaged service with pricing and support from Block or Kyutai aimed at your use case.

The fairest summary: Moshi proved full-duplex speech-to-speech could be open and real. Dessa asks whether removing text entirely makes it more human. The answer will come from independent testing, not launch pages — and this blog will cover those tests when safety researchers publish them.

Sources: Block's official Dessa page (block.xyz/dessa); Kyutai's Moshi paper and project materials (arXiv:2410.00037); The Verge and TechCrunch, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗