Dessa vs Moshi

Two answers to the same problem

Both models attack the same weakness of voice assistants: the trip through text. Instead of chaining speech-to-text, a language model and text-to-speech, each one takes audio in, produces audio out and keeps listening while it talks.

The difference is how much text they use. Moshi starts from a text language model and keeps a text stream inside. Block says Dessa never saw text at all.

Sources: Kyutai on arXiv, Block

Moshi: open and documented

Kyutai, an open-science AI lab in Paris, showed Moshi on 2024-07-03. It released the weights and code on 2024-09-18, a day after its technical report appeared on arXiv. Kyutai describes Moshi as the first real-time full-duplex spoken language model.

Moshi models two audio streams at once, its own voice and the user's. It also predicts the text of its own speech, which Kyutai calls an inner monologue and says improves quality. Its Mimi codec compresses 24 kHz audio into 12.5 frames per second at 1.1 kbps.

Kyutai's model card is clear about the limits: English only, one voice per model, limited ability on complex tasks and no tool use.

Sources: Kyutai, Fortune, Kyutai, Kyutai on arXiv, Kyutai on GitHub, Kyutai on Hugging Face

Dessa: speech only, not released

Block's lab says it trained Dessa from scratch on human conversation, with no text and no pretrained models, on a single NVIDIA GB200 NVL72. According to Block, Dessa runs faster than real time on a MacBook Pro with an M4 Max chip or better.

Block has published recordings, not documentation. There is no paper, model card, parameter count, language list or latency figure, so none of these claims can be checked yet.

Sources: Block

Training data is the real difference

Moshi leans on text twice. Its backbone, Helium, is a 7B text model trained on 2.1 trillion tokens of English text. The speech model then learned from about 7 million hours of mostly English audio transcribed with Whisper, plus 2,000 hours of Fisher phone calls and 20,000 hours of synthetic dialogue. Kyutai's model card says training ran on 1,016 NVIDIA H100 GPUs.

If Block is right, Dessa skips all of that. Block says it learned language from hundreds of times fewer words than a large language model needs, and that one GB200 NVL72, a single rack with 72 Blackwell GPUs, was enough to train it. The two setups are hard to compare, since Block has not said how long training took or how much audio it used.

A speech-only model could learn from talk that has never been written down. It also raises a question Block lists itself: how well such a model learns facts.

Sources: Kyutai on arXiv, Kyutai on Hugging Face, Block, NVIDIA

Running them

You can run Moshi today. Kyutai ships a PyTorch build for research, an MLX build for Macs and iPhones, and a Rust build for production with CUDA or Metal. The PyTorch build needs a GPU with about 24 GB of memory.

The weights are under CC-BY 4.0, which allows commercial use with attribution. Kyutai's model card adds that the model is meant for research and should not give advice or do professional work, so test it carefully before customers hear it.

Kyutai keeps working on the line. On 2026-06-10 it published a version of Moshi post-trained with reinforcement learning, which improves pause handling, turn-taking and backchannels. That checkpoint uses a non-commercial CC BY-NC 4.0 license.

You cannot run Dessa. If Block releases it, the laptop claim would make local use realistic, but the license will decide whether anyone else may host it.

Sources: Kyutai on GitHub, Creative Commons, Kyutai on Hugging Face, Kyutai, Kyutai on Hugging Face, Block

Voices, emotion and safety

Moshi comes with one synthetic voice per model: a male voice in Moshiko and a female voice in Moshika. Kyutai says it trained the model to produce only one voice to avoid impersonation, and its model card rules out impersonation and any other malicious use.

Block says Dessa picks its own voice and personality in each conversation, and that it hears and expresses emotion, sarcasm and laughter. Block has not said whether developers will be able to choose or clone a voice. Its first safety step is early access for a small group of AI safety researchers, who are asked to find risks the team has not anticipated.

Sources: Kyutai on GitHub, Kyutai on Hugging Face, Block