Moshi is a speech-text foundation model and full-duplex spoken dialogue framework from Kyutai, the Paris-based open AI research lab. Unlike voice assistants that chain speech recognition, a language model, and speech synthesis in sequence, Moshi models the conversation as a single audio stream — so it can listen and speak at the same time, react instantly, and handle interruptions the way humans do.
At Moshi’s core is Mimi, a state-of-the-art streaming neural audio codec that compresses speech into tokens the model can predict in real time. The result is remarkably low latency — around 160 milliseconds in theory and roughly 200 milliseconds in practice — which makes conversations feel fluid rather than walkie-talkie-like. Moshi also models an “inner monologue” text stream alongside audio, improving the coherence and factual quality of what it says. The repository ships the full stack in Python and Rust variants, plus an MLX version for Apple silicon. It has more than 11,200 GitHub stars, remains in active development with updates in September 2026, and its code is Apache-2.0 licensed.
Moshi is free and open source, with code under the Apache-2.0 license and model weights published openly by Kyutai for research and development use.
Voice AI researchers and product teams building real-time conversational agents who need natural turn-taking and minimal response delay.