MOSS-LIVE

BetaRealtime interaction

MOSS-Live 1.0 is a full-duplex voice model built for natural, expressive realtime conversation. It listens while it speaks, handles natural interruptions, works with background agents, and can interact over realtime video streams.

View API reference

ENDPOINT

WebSocket /v1/realtime

Current request path

Input

Realtime audio + realtime video stream

Content accepted by the request

Output

Audio + text

Content returned on success

Model overview

MOSS-Live 1.0 is a full-duplex voice model built for natural, expressive realtime conversation. It listens while it speaks, supports natural interruptions, and delegates complex reasoning and long-running tasks to the background for asynchronous execution, continuing to talk with the user while they run. It also accepts realtime video streams, understands the conversation through both sound and image, and streams back voice and text.

Billing

Free for a limited time.

0 credits View full pricing

Modality support

TTextOutput only

AAudioInput and output

IImageNot supported

VVideoInput only · Realtime video stream

Supported endpoints

Single-speaker speech/v1/audio/speechNot supported

Multi-speaker dialogue speech/v1/audio/speech/speakersNot supported

Audio transcription/v1/audio/transcriptionsNot supported

Voice design/v1/audio/voice/generationsNot supported

Multimodal understanding/v1/responsesNot supported

Realtime interaction/v1/realtimeSupported

Image generation/v1/images/generationsNot supported

Capabilities

PerformanceHigh

Inference speedRealtime interaction

Streaming Supported

Async task API Not supported

Semantic interruption and end-of-turn detection
Asynchronous delegation of long-running tasks
Proactive interaction and backchanneling
Ignoring non-addressed speech in multi-party conversations
Realtime audio-visual understanding
Tool calling
Simultaneous interpretation
Playback progress awareness

Versions and snapshots

Model IDmoss-live-1.0