What is Moshi?
Moshi is Kyutai's speech-to-speech dialogue model and the code to run it. Unlike a pipeline that transcribes, thinks and then speaks, Moshi listens and talks at the same time, modelling your audio stream and its own together with a 7B-parameter transformer and a parallel text stream it calls an inner monologue. Its Mimi codec compresses 24 kHz audio to 12.5 frames a second at 1.1 kbps.
The repository has three inference stacks: PyTorch for research, MLX for Macs and iPhones, and Rust with CUDA or Metal for production, plus the web client behind the moshi.chat demo. It ships two voices, Moshiko and Moshika, and the same code runs Kyutai's Hibiki speech translation and its streaming speech-to-text and text-to-speech models.
The Python code and web client are MIT-licensed and the Rust backend Apache 2.0; the weights are CC-BY 4.0. The repository has about 11,500 GitHub stars, the moshi package reached 0.2.13 on PyPI in February 2026, and the last commit landed on 16 May 2026.
Is Moshi worth it?
Moshi's latency is the reason to look at it. Kyutai puts the theoretical floor at 160 ms (an 80 ms Mimi frame plus 80 ms of acoustic delay) and measures about 200 ms end to end on an L4 GPU. Because it models both speakers at once, it can be interrupted and can speak while you speak, which pipelines that chain speech recognition, a language model and speech synthesis have to imitate with voice-activity detection.
The model behind it has not moved since 2024. It speaks English only, the FAQ says changing its voice or personality would need fine-tuning, and the pre-training data will not be released. The MLX and Rust versions stop after about 5 minutes of conversation because they use a fixed buffer; the PyTorch version keeps going, but Kyutai says that is mostly untested and quality is likely to degrade.
Hardware is the practical barrier. The PyTorch stack has no quantisation and wants a 24 GB GPU; the Rust backend runs int8 on CUDA, a 12 GB card works only with a workaround from the issue tracker, and 8 GB is not possible. On a Mac, MLX with 4-bit weights is the route; Kyutai reports that going below 4 bits ruins quality and does not expect it to run on an M1.
Deployment defaults need a careful read. The PyTorch server listens on localhost port 8998 with no authentication in the code, while the Rust backend's sample config binds to 0.0.0.0 with a self-signed certificate, reachable from the network. The --gradio-tunnel option publishes a public URL routed through the US that Kyutai says can add up to 500 ms from Europe. A production voice agent will answer better with a pipeline around a stronger language model; Moshi is the reference for full-duplex research and for anyone building the next one.
Where it earns its place
A speech researcher at a university lab
Study turn-taking and interruptions in spoken dialogue
Run the PyTorch server on a 24 GB GPU and log both audio streams alongside the inner-monologue text Moshi predicts. The open weights and the arXiv technical report give a citable baseline, and the code is free under MIT.
A voice-interface engineer at a consumer-robotics start-up
Feel what sub-second, interruptible conversation does to a product before building one
Build the Rust backend with CUDA, start it with config-q8.json for int8 weights and open the web UI on port 8998. Treat it as a prototype: English only, a fixed persona and sessions that end after about 5 minutes.
A Mac developer exploring on-device voice
Try a speech-to-speech model without a cloud bill
pip install moshi_mlx on Python 3.12 and run python -m moshi_mlx.local -q 4 on Apple silicon. Kyutai tested on an M3 MacBook Pro; the command-line client has no echo cancellation, so use the web UI or headphones.
Before you commit
- Setup
- pip install moshi (Python 3.10 or later, 3.12 recommended), then python -m moshi.server downloads weights from Hugging Face and serves a web UI at localhost:8998. The Rust backend needs the Rust toolchain plus CUDA with nvcc, or Metal on macOS; Docker Compose works on CUDA.
- Integrations
- A web UI, Python and Rust command-line clients and a Gradio demo share one streaming protocol; rustymimi provides Python bindings to the Mimi codec; the same stack runs Hibiki and Kyutai's streaming speech-to-text and text-to-speech models.
- Data and privacy
- Self-hosted, so audio stays on your machine unless you enable --gradio-tunnel, which exposes the server through a public URL routed via the US. The Rust sample config binds 0.0.0.0:8998 with a self-signed certificate, and the PyTorch server has no login. Weights are CC-BY 4.0, which requires attribution.
- Cost at scale
- Code and weights are free; the cost is GPU time, from a 24 GB card for PyTorch or an int8 Rust deployment on CUDA. Kyutai does not publish how many simultaneous conversations one GPU handles.
- Limits
- English only, 24 GB of VRAM for PyTorch, sessions of about 5 minutes on MLX and Rust, no official Windows support, and no release of the pre-training dataset.
What’s good
- About 200 ms end-to-end latency on an L4 GPU
- Full duplex: listens while it speaks and handles interruptions
- PyTorch, MLX and Rust stacks plus a ready web client
- Open weights under CC-BY 4.0; code under MIT and Apache 2.0
What’s not
- English only; a new voice or persona needs fine-tuning
- 24 GB GPU for PyTorch; 8 GB cards cannot run it
- MLX and Rust sessions stop after about 5 minutes
- The model is unchanged since its 2024 release
Best for
Speech researchers and voice-interface engineers who want to study or extend a full-duplex spoken dialogue model on their own GPUs.
Moshi alternatives
Moshi: common questions
- Is Moshi worth it?
- Moshi still shows what a truly conversational voice model feels like, but a 2024 English-only model that wants a 24 GB GPU is a research platform, not a voice agent to ship. We rate it 3.6 out of 5.
- How much does Moshi cost?
- Free: code under MIT and Apache 2.0, weights under CC-BY 4.0. No paid edition.
- What are the downsides of Moshi?
- English only; a new voice or persona needs fine-tuning; 24 GB GPU for PyTorch; 8 GB cards cannot run it; mLX and Rust sessions stop after about 5 minutes; the model is unchanged since its 2024 release.
- Who is Moshi best for?
- Speech researchers and voice-interface engineers who want to study or extend a full-duplex spoken dialogue model on their own GPUs.
- How much does Moshi cost for a team or in production?
- Code and weights are free; the cost is GPU time, from a 24 GB card for PyTorch or an int8 Rust deployment on CUDA. Kyutai does not publish how many simultaneous conversations one GPU handles.
- Is Moshi safe for company data?
- Self-hosted, so audio stays on your machine unless you enable --gradio-tunnel, which exposes the server through a public URL routed via the US. The Rust sample config binds 0.0.0.0:8998 with a self-signed certificate, and the PyTorch server has no login. Weights are CC-BY 4.0, which requires attribution.
- What are the best Moshi alternatives?
- The closest alternatives we have reviewed are VibeVoice and audio.cpp.
Answers checked Nov 14, 2026.