Kyutai's open-weights, full-duplex speech model that listens and talks at the same time, with a free browser demo.
Moshi demo page
Overview
What is Moshi?
The open full-duplex voice model you can talk to and interrupt.
Moshi is an open-weights, full-duplex speech model from Kyutai, a Paris-based open-science AI lab. Moshi listens and speaks at the same time, generating audio directly, and it can be interrupted mid-sentence. Unlike Block's Dessa, which claims to have learned from speech alone, Moshi was built on top of a text model.
Kyutai released Moshi's model weights and inference code, with backends for PyTorch, MLX for Apple Silicon, and Rust. A free browser demo at moshi-chat.kyutai.org lets visitors try a live conversation, with demo conversations limited to 5 minutes as shown on the demo page. Moshi also has a vision variant, MoshiVis, for talking about images in real time.
Platforms and languages
Moshi Availability
Platforms
Languages
Capabilities
Moshi Key Features
Full-duplex speech dialogue
Listens and speaks at the same time, with interruptions handled mid-sentence.
Open weights and code
Model weights and inference code are openly released by Kyutai.
Free browser demo
Try a live conversation in the browser demo, with sessions limited to 5 minutes.
Local and on-device runs
Run Moshi locally on a GPU, or on Apple Silicon with the MLX backend.
Vision variant MoshiVis
MoshiVis extends Moshi to discuss images in real time.
Best for
Who uses Moshi?
Developers and researchers
build and study real-time, full-duplex voice assistants with open weights
Voice AI enthusiasts
try a live full-duplex conversation in the browser demo
Plans and access
Moshi Pricing
Free
Common questions
Moshi FAQs
How do I try Moshi?
Open moshi-chat.kyutai.org in a browser, press Connect, and talk. Demo conversations are limited to 5 minutes, and Chrome works best according to the demo page.
Is Moshi open source?
Yes. Kyutai released Moshi's model weights and inference code, so developers can run it locally with a capable GPU or on Apple Silicon via the MLX backend.
What does full-duplex mean for Moshi?
Moshi listens and speaks at the same time, while most voice assistants wait for you to finish before replying. That full-duplex design lets you interrupt Moshi mid-sentence.
How is Moshi different from Dessa?
Moshi was built on top of a text model, while Block says Dessa learned from speech alone. Moshi is therefore the closest open comparison point for Dessa's claims.
Reviews
See what the community thinks and share your experience.
—
Based on 0 ratings
Rating distribution
Community reviews
No reviews
No written reviews yet.
