
Nonprofit · Paris, France
Kyutai
Paris non-profit open-science lab building speech-native and multimodal models, creator of Moshi, Mimi, Unmute and Helium.
Paris, France
- Founded
- 2023
- Type
- Nonprofit
- Based in
- Paris, France
- Key papers
- 3
- Featured projects
- 8
About
Kyutai is an open-science AI research lab in Paris, funded by the Iliad Group, the CMA CGM Group and Schmidt Sciences. Its team is led by Patrick Pérez as chief executive, with Alexandre Défossez as head of research and Laurent Mazaré as head of engineering. Edouard Grave and Hervé Jegou are listed among its alumni.
The lab is known for Moshi, described as the first speech-native dialogue system, which processes speech directly rather than converting it to text and back. It also released Unmute, which gives any text model real-time listening and speaking, and Helium, a multilingual two-billion-parameter language model. Papers, code and weights are published openly.
Its recent research output covers text-to-speech, video autoencoders and world models.
Mission and ambitionsKyutai aims to advance open research on speech and multimodal AI, publishing models, code and papers for the community.
Known for
- 01
Moshi speech-native dialogue model
- 02
Unmute streaming speech interface
- 03
Helium multilingual language model
Milestones
- 2023
Lab launched
Kyutai was announced in Paris as a non-profit open-science lab.
- 2024
Moshi
Released a real-time speech-native dialogue model with open weights.
- 2024
Mimi codec
Published the streaming neural audio codec that underpins Moshi.
- 2025
Unmute
Released a system adding real-time speech input and output to text models.
- 2025
Helium 1
Released a modular multilingual language model of 2B parameters.
Open source
Moshi is a speech-text foundation model and full-duplex spoken dialogue framework. It uses Mimi, a state-of-th (11,542 stars)
A TTS that fits in your CPU (and pocket) (10,320 stars)
Kyutai's Speech-To-Text and Text-To-Speech models based on the Delayed Streams Modeling framework. (3,053 stars)
Make text LLMs listen and speak (1,553 stars)
Model: 308,499 downloads, 266 likes.
Model: 126,722 downloads, 384 likes.
- kyutai/DailyTalkContiguousdataset
Dataset: 2,646 downloads, 22 likes.
Space: 56 likes on the Hub.
Research papers
Notable releases
Openness
Models and code are released openly, mostly under Apache-2.0 or MIT licences, with technical papers.
People
- Patrick PérezChief Executive Officer
- Alexandre DéfossezHead of Research
- Laurent MazaréHead of Engineering
- Neil ZeghidourAudio Research Advisor
- Václav VolhejnTechnical staff
- Amélie RoyerTechnical staff
Listed from the lab's own pages and public profiles.
Contact and community
Contact address printed on the website.
Sources
Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.



