Kyutai · siet September 2024 (technischer Bericht, arXiv 17.09.2024; Demo-Launch 3. Juli 2024) · 2× · tolest 27. Okt. 2026
25
Momentum
Moshi is a speech-text foundation model and full-duplex spoken dialogue framework from Kyutai, a non-profit research lab in Paris. It processes speech directly (speech-to-speech) and models its own and the user's audio streams in parallel. According to the model card, it supports English; the weights are under CC-BY 4.0, the code under MIT and Apache-2.0.
Momentum-Verloop
30.07.28.10.
Features
| Real-Time Streaming | Yes: full-duplex, speech in and out simultaneously |
| Latency | Theoretical 160 ms, ~200 ms in practice |
| License | Weights CC-BY 4.0; code MIT (Python) / Apache-2.0 (Rust) |
| Platform | Python (PyTorch, MLX), Rust (Candle); runs locally, ~24 GB GPU for PyTorch |
| Price | Free (open weights; self-hosting costs only hardware) |
| Release Date | July 2024 demo; September 2024 model/paper (v0.1) |
| Languages | English (per model card) |