Moshi

Kyutai is an AI research lab based in Paris. Its stated mission is to build and democratize artificial general intelligence through open science. The lab works across audio, vision and language, and says it aims to build efficient models that combine several input types and to release the tools needed to use and understand them.

In audio, Moshi is described as the first speech-native dialogue system. It processes speech directly instead of converting it to text and back, which the page says gives minimal latency and lets it understand emotions and other non-verbal cues. Unmute gives any text LLM real-time voice by pairing it with streaming text-to-speech and speech-to-text models. Hibiki-Zero translates speech to speech in real time, and Muscriptor transcribes music into multi-instrument MIDI.

In vision, the lab explores video and model efficiency. MoshiVis and CASA use cross-attention to combine visual input with text or speech for real-time interaction with video. MIRA is a real-time multiplayer world model, built with General Intuition and in collaboration with Epic Games, trained on 10k hours of Rocket League games.

In language, Helium 1 is a modular, multilingual 2B-parameter LLM, and the lab says it is releasing the codebase and tools needed to reproduce its training dataset. Pocket TTS, a 100M-parameter multilingual text-to-speech model, is open source and runs faster than real time on CPU.

Kyutai makes its research available through scientific publications, tutorials, online demos and open-source releases, with code on GitHub and models on Hugging Face. The page says the lab is funded by Iliad Group, CMA CGM Group and Schmidt Sciences. It does not list prices. Most projects are presented as research releases with free online demos.