Moshi

Moshi is an open real-time voice AI model developed by the Paris lab Kyutai and open-sourced on 18 September 2024 (first demo in July 2024). It is a speech-to-speech dialogue system that processes speech directly, without an intermediate text conversion, and operates in full duplex — listening and speaking at the same time.

Its architecture combines the Helium language backbone (7B parameters, trained on 2.1T tokens), the Mimi streaming neural audio codec (12 Hz, 1.1 kbps) and the “Inner Monologue” method, in which the model predicts text tokens as a prefix to audio tokens, improving the linguistic quality of speech. The model has about 8 billion parameters.

Moshi achieves a theoretical latency of 160 ms (about 200 ms in practice), enabling natural real-time conversation; it also understands emotions and non-verbal aspects of communication. It takes audio (24 kHz, mono) as input and produces speech plus time-aligned text as output. It was trained on about 7 million hours of audio and 20,000 hours of synthetic dialogue.

The model is open: weights under CC-BY-4.0 and code under Apache-2.0 (GitHub, Hugging Face). Variants include Moshiko (male voice) and Moshika (female voice). Kyutai is a non-profit AI lab based in Paris, funded by Iliad, CMA CGM and Schmidt Sciences, among others.