Moshi

Moshi, from Kyutai Labs, is a speech-text foundation model built for full-duplex spoken dialogue: it models the user's audio and its own audio as separate parallel streams, so it can listen and speak at the same time instead of taking turns. Architecturally it pairs a 7B parameter Temporal Transformer with a smaller Depth Transformer that handles inter-codebook structure, reaching a theoretical latency of 160 ms and about 200 ms in practice on an L4 GPU. Audio flows through Mimi, the project's streaming neural codec, which represents 24 kHz audio at 12.5 Hz frames and 1.1 kbps with 80 ms latency, a frame rate close to text token rates that keeps autoregressive steps manageable. The repository ships three backends: PyTorch for research, MLX for Apple silicon, and Rust with Candle for production serving. Code is released under MIT (Python) and Apache (Rust) licenses with model weights under CC-BY 4.0, making it a key open reference implementation for real-time voice AI.