model · Kyutai (French AI lab)

Moshi (Kyutai)

Moshi, from French lab Kyutai, is a genuinely different architecture from everything else on this page: instead of chaining separate speech-to-text, language model, and text-to-speech stages, it's one model trained end-to-end on audio that listens and speaks simultaneously (full-duplex), the way humans actually converse — it can be interrupted mid-word and react instantly because there's no hand-off between separate components. That collapses latency to under 200ms and makes overlapping speech and natural back-channeling ('yeah', 'right') possible in a way cascaded pipelines struggle to fake. It's released under CC-BY-4.0 and can be self-hosted entirely in the EU, which matters if data residency or privacy is a concern. Ultravox is the other notable open speech-native model, taking a slightly different approach (audio feeds directly into an LLM, skipping a separate ASR step). Both require a decent GPU to run at real-time speed — this is the frontier-research end of open source, not a weekend project.

• free
• self-host
• Open source