Moshi: a speech-text foundation model for real-time dialogue research paper summary (October 2026)

The Moshi paper from the French research lab Kyutai introduces a speech-to-speech foundation model for real-time spoken dialogue. The open repository offers inference in PyTorch, MLX for Macs and iPhones, and Rust for production.

Classic voice assistants chain voice activity detection, speech recognition, a text model and speech synthesis, which adds seconds of delay and drops emotion and overlapping speech. Moshi models the spoken dialogue directly, predicting its own speech, the user's speech and an inner text monologue with a 7-billion-parameter temporal Transformer.

The Mimi codec compresses 24 kHz audio to 12.5 Hz at 1.1 kbps with 80 ms latency. The same codebase runs Kyutai's Hibiki speech translation and its text-to-speech and speech-to-text models. The Python code is under MIT, the Rust backend under Apache 2.0 and the model weights under CC-BY 4.0. The repository had 11,158 stars and was updated on September 9, 2026.

• Enables natural voice agentsBecause Moshi listens while it speaks, it can handle interruptions and quick back-and-forth. Customer service or internal help desk teams can prototype voice agents that feel less like a phone menu.
• Runs speech models on deviceThe MLX implementation runs inference on Macs and iPhones. Product teams can test voice features locally without streaming audio to a cloud service.
• Compresses speech efficientlyThe Mimi codec encodes speech at 1.1 kbps while streaming. Developers building voice pipelines can reuse it for low-bandwidth audio transmission or as a tokenizer for their own models.

• European research lab based in France; code under MIT and Apache 2.0, model weights under CC-BY 4.0.
• Three implementations: PyTorch for research, MLX for Apple devices, Rust for production.
• Speech-to-speech design without a separate recognition and synthesis pipeline.

The paper is freely available on arXiv, and the accompanying code runs on your own hardware in any country.

Moshi is a research model. Test language coverage and answer quality for your use case before putting it in front of customers. Voice recordings are personal data, so decide where inference runs and how audio is stored. The model weights are under CC-BY 4.0, which requires attribution when you redistribute or adapt them.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews