Kyutai
Paris open-science lab behind Moshi, the open real-time voice AI.
About
Kyutai is a Paris-based non-profit AI research lab founded in 2023 by French AI researchers and entrepreneurs. It focuses on open-source, privacy-centric AI for speech, text, and multimodal applications, with a philosophy of privacy by design and on-device processing that aligns with GDPR and the EU AI Act.
Its flagship project is Moshi, an open-source full-duplex speech-text foundation model that can listen and speak simultaneously with sub-300ms latency, handling interruptions and overlapping speech like a real conversation. The Moshi stack includes the Mimi streaming neural audio codec and ships in PyTorch, MLX, and Rust implementations, plus a live demo at moshi.chat.
Kyutai has also released Unmute, a modular framework that gives any text-based LLM a real-time voice via streaming STT, an LLM, and streaming TTS, and Hibiki, a simultaneous speech translation model. Its open weights have made Moshi a foundation for projects such as Nvidia's PersonaPlex.
What it can do
- Moshi full-duplex spoken dialogue model
- Mimi streaming neural audio codec
- Unmute real-time voice framework for any LLM
- Hibiki simultaneous speech translation
- Permissive open-source releases
Setup
Clone the kyutai-labs/moshi repository or try the live demo at moshi.chat.
Prerequisites
For local inference, a GPU or Apple Silicon Mac; a browser for the demo.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 133; Trending score 80/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.