pip install moshi

git clone https://github.com/kyutai-labs/moshi
cd moshi
pip install -e .

docker pull kyutai/moshi:latest
docker run --gpus all -p 8998:8998 kyutai/moshi:latest

from moshi import Moshi
 
# Initialize model
model = Moshi.from_pretrained("kyutai/moshi-7b")
 
# Create a session
session = model.create_session()
 
# Process audio and get response
response = session.process(input_audio)

import asyncio
from moshi import Moshi, AudioIO
 
async def run_conversation():
    model = Moshi.from_pretrained("kyutai/moshi-7b")
    session = model.create_session()
 
    # Bidirectional audio streaming
    audio_io = AudioIO()
 
    async with session.stream() as stream:
        async for input_chunk in audio_io.input():
            response_chunk = await stream.process(input_chunk)
            await audio_io.output(response_chunk)
 
asyncio.run(run_conversation())

session = model.create_session(
    # Full-duplex settings
    duplex_mode="full",          # "full", "half", or "adaptive"
    overlap_handling="natural",   # How to handle overlapping speech
 
    # Timing controls
    min_response_gap=50,         # ms before responding
    barge_in_sensitivity=0.7,    # How easily user can interrupt
)

session = model.create_session(
    backchanneling=True,  # Enable "mhm", "I see", etc.
    backchannel_frequency="natural"  # or "minimal", "frequent"
)

# Available voices
voices = model.list_voices()
# ['moshi-default', 'moshi-calm', 'moshi-energetic']
 
session = model.create_session(voice="moshi-calm")

session = model.create_session(
    language="en",  # or "fr"
    accent="neutral"  # or "british", "american", "french"
)

session = model.create_session(
    system_prompt="""You are a friendly language tutor helping users practice French.
    Speak slowly and clearly. Gently correct pronunciation mistakes.
    Encourage the user and keep the conversation flowing naturally."""
)

from moshi import Moshi, WebSocketServer
 
model = Moshi.from_pretrained("kyutai/moshi-7b")
 
server = WebSocketServer(
    model=model,
    host="0.0.0.0",
    port=8998
)
 
server.run()

const ws = new WebSocket('ws://localhost:8998/ws');
 
// Set up audio context
const audioContext = new AudioContext({ sampleRate: 24000 });
const processor = audioContext.createScriptProcessor(2048, 1, 1);
 
// Send audio
processor.onaudioprocess = (e) => {
    const audioData = e.inputBuffer.getChannelData(0);
    ws.send(audioData.buffer);
};
 
// Receive audio
ws.onmessage = async (event) => {
    const audioData = await event.data.arrayBuffer();
    playAudio(audioData);
};

version: '3.8'
services:
  moshi:
    image: kyutai/moshi:latest
    runtime: nvidia
    ports:
      - "8998:8998"
    environment:
      - CUDA_VISIBLE_DEVICES=0
      - MODEL_ID=kyutai/moshi-7b
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

apiVersion: apps/v1
kind: Deployment
metadata:
  name: moshi
spec:
  replicas: 2
  selector:
    matchLabels:
      app: moshi
  template:
    spec:
      containers:
      - name: moshi
        image: kyutai/moshi:latest
        resources:
          limits:
            nvidia.com/gpu: 1
        ports:
        - containerPort: 8998

model = Moshi.from_pretrained(
    "kyutai/moshi-7b",
    quantization="int8"  # Reduces VRAM by ~50%
)

model = Moshi.from_pretrained(
    "kyutai/moshi-7b",
    max_batch_size=4,  # Handle 4 conversations concurrently
    dynamic_batching=True
)