Kyutai

Product overview

Kyutai is a non-profit open-science AI lab based in Paris, established in November 2023 with a €300 million endowment from French billionaire Xavier Niel (Iliad/Free), CMA CGM CEO Rodolphe Saadé, and former Google CEO Eric Schmidt. Led by Patrick Pérez with Yann LeCun among its scientific advisors, the lab operates from Station F and uses approximately 1,000 NVIDIA H100 GPUs provided at cost by Scaleway, Niel's cloud company. Kyutai's first major release was Moshi, a 7-billion-parameter speech-native dialogue model capable of full-duplex conversation with 200-millisecond latency and over 70 emotional styles. Moshi and its Mimi neural audio codec were fully open-sourced in September 2024. Helium-1 is a 2-billion-parameter multilingual model designed for mobile devices, covering six European languages. Hibiki-Zero provides end-to-end real-time speech translation. The lab also released Kyutai TTS (including a 100-million-parameter pocket version), speech-to-text models, and MoshiVis for image understanding. All models are released as open source with weights, training code, and datasets freely available on GitHub and Hugging Face. Gradium, a commercial spin-off, packages the research into production-ready voice systems. The lab's small team developed Moshi before OpenAI shipped its comparable GPT-4o voice mode. KEY FEATURES: - All models fully open-source with weights, training code, and datasets - Moshi: 7B-parameter speech dialogue model with 200ms latency and full-duplex conversation - Helium-1: 2B multilingual model for mobile covering six European languages - Runs on Scaleway (French cloud) infrastructure with NVIDIA H100 GPUs - Gradium commercial spin-off for production voice AI deployments

Sovereignty