Official announcements or releases from Kyutai Labs

French AI Lab Kyutai Unveils Moshi: A Real-Time Multimodal Model with 160ms Latency

Jul 3, 2024, 03:55 PM

Kyutai Labs, a French AI lab with $300M in funding, has unveiled Moshi, a groundbreaking real-time multimodal model capable of listening and speaking. This model, which runs on Apple laptops and consumer GPUs, boasts a low latency of 160ms and can function without internet access, ensuring no data leaves the user's computer. Developed by an eight-person team over six months, Moshi is set to be open-sourced, allowing for further enhancements. It features a 7B Multimodal LM with 2 channel I/O and a Real-Time Factor of 2. Designed for smooth, natural, and expressive communication, Moshi can serve as a coach, companion, or roleplay character. The model's performance, though slightly robotic, is remarkably fast and efficient, making it a strong competitor to OpenAI's GPT-4o.