kyutai-labs/moshi: Speech-text foundation model and dialogue framework

github.com7 min read

Kyutai Labs has released Moshi, a speech-text foundation model and full-duplex spoken dialogue framework, alongside Mimi, a streaming neural audio codec. Moshi achieves a theoretical latency of 160ms and models two audio streams alongside internal text tokens to improve generation quality. The codebases and pre-trained model weights are made available across PyTorch, MLX, and Rust stacks under open licenses. Open article →