Moshi
Kyutai's open full-duplex speech model: it listens and talks at the same time, with a 160ms theoretical latency.
About
Moshi is the French lab Kyutai's speech-text foundation model for real-time dialogue, released September 2024. It is full-duplex by construction: two audio streams (the user speaking, Moshi speaking) plus an 'inner monologue' of text tokens are modeled together, so interruptions, overlaps, and backchannels work the way they do in human conversation. A 7B temporal transformer sits on the Mimi streaming neural audio codec, with a theoretical latency of 160ms and practical latency as low as 200ms on an L4 GPU.
The weights are CC-BY 4.0 and the code MIT/Apache, with PyTorch, MLX (iPhone/Mac), and Rust stacks plus a live demo at moshi.chat. Moshi became the base that later open voice work builds on: NVIDIA's 2026 PersonaPlex models are explicitly built on the Moshi architecture. There are no standard academic benchmarks for full-duplex dialogue quality, so evaluation here is qualitative and community-driven rather than leaderboard-based.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No standard public benchmarks for full-duplex spoken dialogue; evaluation is qualitative (latency, turn-taking naturalness) rather than leaderboard-based
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: real Kyutai release Sep 2024 with CC-BY-4.0 weights and MIT/Apache code across three inference stacks; peer-reviewed paper (arXiv:2410.00037); became the architectural base for NVIDIA's 2026 PersonaPlex open voice models. Open weights. Strengths: pioneering open full-duplex design, permissive licensing, genuine downstream influence. Docked for: (1) no standard benchmarks exist for its core capability, so quality claims are qualitative; (2) 2024-era speech quality now trails newer cascaded and native voice systems.
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
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Kyutai release Sep 2024; paper arXiv:2410.00037; first major open full-duplex voice model
- · CC-BY-4.0 weights, MIT/Apache code; PyTorch/MLX/Rust stacks; moshi.chat demo
- · Architectural base for NVIDIA PersonaPlex (2026) — real downstream influence
- · No standard benchmarks for full-duplex dialogue; 2024-era speech quality