Kyutai
Kyutai is an independent nonprofit AI research lab in Paris. It develops language, speech and visual models and publishes research code and model releases with terms that vary by project.
Company & leadership
- CEO
- Patrick Pérez, listed by the current team page.
- Research and engineering leads
- Alexandre Défossez leads research; Laurent Mazaré leads engineering.
- Founding organizations
- Iliad Group, CMA CGM Group and Schmidt Futures. The current site names Schmidt Sciences among its donors.
- Launch funding
- Iliad and CMA CGM each committed €100 million. The launch announcement described nearly €300 million in total support; this is not a current valuation.
Founders
- Xavier Niel, Rodolphe Saadé and Eric Schmidt.
The organization is a research lab
Kyutai was founded to conduct open AI research outside a single commercial product line. Its Paris launch brought together support from telecoms, transport and philanthropic organizations. Its models should be assessed by their own releases rather than by assuming one company-wide API or license.
The team page separates current staff from alumni
Patrick Pérez is listed as CEO. The current research and engineering leaders differ from some names in early announcements. The site lists Edouard Grave and Hervé Jégou among alumni, so their earlier leadership titles do not describe current roles.
The lab develops more than speech models
Moshi and Hibiki are prominent speech projects. The wider program includes Helium language models, visual fusion, view generation, robotics and world simulation. Some releases are reusable components, such as encoders and codecs, rather than complete assistants.
Pocket TTS has separate trained language releases
The public code lists language checkpoints, deeper variants and small English models. The english name selects the current October release. Voice-cloning weights require accepting the model page conditions; a separate no-voice-cloning repository provides another access route. Both repositories identify CC BY 4.0. These terms differ from the code’s MIT license.
The smallest speech models reduce several components
The October Pocket TTS release includes 18.2M, 6.9M and 1.9M models, counting their decoders. Distillation reduces the backbone, sampling head and decoder separately. REFL variants add reward fine-tuning. The 1.9M version is experimental, and published quality or speed comparisons depend on the reported setup.
Helium makes data specialization visible
Helium’s specialized checkpoints continue training on selected data classes. They are not separate hosted products. The final model demonstrates how selected trained weights can be combined. Dactory’s classifiers support the data-selection stage.
Collaborations remain credited to their developers
MIRA is developed with General Intuition and Epic Games. MuScriptor is a joint music-transcription project with Mirelo. These collaborations do not make every product offered by those partners a Kyutai model.