# kyutai-labs/moshi

Moshi is a speech-text foundation model and full-duplex spoken dialogue framework. It uses Mimi, a state-of-the-art streaming neural audio codec.

Repository: https://github.com/kyutai-labs/moshi
Canonical: https://ross.abutalabs.com/products/moshi
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-05-16T13:14:41+00:00

## Health v2 (maintenance only)
Score: 62/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 82, release rhythm 40, longevity 54
- inputs: {"age_days": 756, "days_push": 109, "days_rel": 710, "gap_med": 2, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 10949, forks 1017 (observed 2026-08-28T04:10:44.812280+00:00)

## What it is
Moshi is a speech-text foundation model and full-duplex spoken dialogue framework from Kyutai, built around the Mimi streaming neural audio codec. It ships three inference stacks (PyTorch for research, MLX for on-device Mac/iPhone, Rust for production) and achieves real-time conversational latency around 200ms.

## Use cases
- build a real-time voice chatbot that can talk and listen simultaneously
- run a spoken dialogue model locally on a Mac with MLX
- deploy low-latency speech-to-speech conversation on a GPU server
- experiment with streaming neural audio codecs like Mimi
- create a voice assistant with natural full-duplex conversation
- integrate speech-text foundation model inference into a Rust application

## When to choose
- you need real-time, full-duplex spoken dialogue rather than turn-based speech recognition plus TTS
- you want on-device voice AI on Apple silicon via the MLX stack
- you need a production-grade low-latency voice model with a Rust inference server
- you are researching speech-text foundation models and streaming audio codecs

## When to avoid
- you only need simple batch speech-to-text or text-to-speech without real-time dialogue
- you have no GPU or Apple silicon and need CPU-only inference
- you need extensive fine-tuning tooling out of the box (fine-tuning lives in a separate repo)
- you need a mature managed API with SLAs rather than self-hosted open-source models

## Facets
- artifact type: library
- maturity: active
- function: speech-recognition, tts, llm-inference, machine-learning, audio-processing
- domain: artificial-intelligence, speech-processing, large-language-models
- platform: python, rust
- tags: full-duplex-dialogue, neural-audio-codec, streaming-inference, voice-assistant, foundation-model, mlx, audio, macos, gpu, web-server

## Member repositories
- kyutai-labs/moshi (main) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:44.812280+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:17:18.468161+00:00, confidence not recorded.
  - readme: https://github.com/kyutai-labs/moshi (fetched 2026-08-28T04:10:44.812280+00:00, sha 04456364989a)
- Data as of 2026-08-30T08:39:29.467469+00:00.