TOOLORA / Local reading edition
Moshi
Model user and assistant audio streams together for full-duplex spoken dialogue, with PyTorch, MLX and Rust inference implementations.
Overview
Moshi jointly models user audio, system audio and text, allowing listening and speaking to overlap instead of waiting for a complete recording to be transcribed. It uses the Mimi streaming codec, with PyTorch research code, Apple-device MLX, a Rust backend and web client. Hardware and quantization vary by implementation. Python and client code use MIT, Rust uses Apache, and official weights use CC-BY 4.0.
- Company country / region
- France · Company headquarters
Official sources and references
- Official websitehttps://github.com/kyutai-labs/moshi
Descriptions summarize public information; consult official websites for features, pricing and regional availability.