Moshi
Kyutai's full-duplex speech-to-speech dialogue model — it listens and talks at the same time.
About
Moshi is a speech-text foundation model from Paris nonprofit lab Kyutai for real-time spoken dialogue. It uses an RQ-Transformer (a 7B Helium temporal transformer plus a depth transformer) over Mimi, a streaming neural audio codec, modeling the user's and its own audio streams in parallel with an inner text monologue — so it handles interruptions and overlapping speech like a phone call instead of a turn-based pipeline, at around 160-200ms latency.
The model weights are CC-BY 4.0 with code under MIT/Apache 2.0, covering commercial use. It is genuinely built for two-way conversation rather than dictation or transcription, and as a research release it can be less polished than commercial voice APIs on noise handling and long conversations.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Reference open full-duplex speech model; basis for NVIDIA PersonaPlex; Weights CC-BY 4.0, code MIT/Apache 2.0; Research release; less polished than commercial voice APIs on noise and long conversations. License: Weights CC-BY 4.0; code MIT and Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Reference open full-duplex speech model; basis for NVIDIA PersonaPlex
- · Weights CC-BY 4.0, code MIT/Apache 2.0
- · Research release; less polished than commercial voice APIs on noise and long conversations