← Back to Voice models
Kyutai
Moshi
Released Sep 18, 2024
Kyutai's open-weight full-duplex spoken dialogue model, which models the user's and its own speech as parallel streams. It's the base that NVIDIA's PersonaPlex builds on.
- Price per minute
- —
- Pricing
- Open weights; free online demo at moshi-chat.kyutai.org.
- Inputs
- audio
- Tool calling
- No
- Languages
- 1
- Voices
- 2
- Big Bench Audio
- 4.4%
- Access
- Generally available
- Licence
- Creative Commons Attribution 4.0
Strengths & weaknesses
Strengths
- About 200 ms practical latency on an L4 GPU (160 ms theoretical)
- Weights CC-BY 4.0; runs on PyTorch, MLX and Rust/Candle
- Inner-monologue text tokens alongside speech
Weaknesses
- Big Bench Audio 4.4%, last on AA
- No tool access; English only
- Two fixed synthetic voices (Moshiko, Moshika)
Not notable: last on AA Big Bench Audio (#44, 4.4%). releaseDate is the open-weights release, from the Kyutai blog URL (/blog/2024-09-18-moshi-release); Kyutai first demoed Moshi in July 2024. Code licences: Python MIT, Rust Apache. 7B temporal transformer (the HF card lists 8B BF16 tensors). Older than the other entries but still available, so drop it if age matters. The HF card says the model is trained to produce one voice per checkpoint; the two released variants give 2 voices.
Links
API model strings
- huggingface
kyutai/moshiko-pytorch-bf16 - huggingface
kyutai/moshika-pytorch-bf16