Alibaba Qwen Team Releases Qwen3.8-LiveTranslate, Real-Time Interpretation Model With 2.3-Second Average Lag

The central technical change is a new Interleave architecture. Qwen reports improvements in faithfulness, fluency and conciseness. Simultaneous interpretation requires a tradeoff: waiting longer gives the model more context, while speaking sooner cuts delay for the listener. LAAL, or Length-Adaptive Average Lagging, measures how far the translation trails the source speech on average and avoids rewarding systems that over-generate output, according to the report.

Qwen3.8-LiveTranslate adds real-time speaker diarization, synchronized bilingual display and long-context disambiguation. Speaker diarization distinguishes speakers in multi-party speech and preserves each speaker’s voice through more stable voice cloning. The API exposes cloning modes, including an always mode that re-clones before each response for multi-speaker sessions. Synchronized bilingual display shows source text and translation on screen together; in the API, source transcription streams as its own events next to the translation stream. Long-context disambiguation uses conversation history to resolve names and terminology, so a name introduced early in a meeting stays consistent later in the translation.

QwenCloud describes the model as the real-time version of Qwen3.8-LiveTranslate-Flash. It is built on the Qwen-Omni stack, large-scale multimodal data, cross-language and cross-modal alignment, and visual enhancement. The Flash model also supports offline audio and video translation.

The model understands 60 languages and can speak 29 of them, returning audio plus text. The remaining 31 return text only. Speech output covers Chinese, English, Arabic, German, French, Spanish, Japanese, Korean, Hindi and others. Inputs are audio and optional images; outputs are text and audio. Visual cues such as lip movements, gestures and on-screen text are intended to help in noisy rooms and with ambiguous words. The documentation recommends sending no more than two images per second. Teams can also set hotwords that map source terms to fixed target translations, with documentation recommending no more than 1,000 hotwords.

Developers connect through the WebSocket Realtime API with the model ID qwen3.8-livetranslate-flash-realtime. The default turn detection type is speaker_detection. Clients stream audio continuously and receive server-generated responses. Default audio is 16 kHz PCM in and 24 kHz PCM out. The default voice is Tina. Users can set session.output_modalities to text only, or text and audio. The report says clients should always send session.finish before closing, or the final segment is lost.

Singapore list pricing per 1 million tokens is $7.50 for audio input, $0.55 for image input, $20 for text output and $30 for audio output. Beijing pricing is lower, at $5.653, $0.466, $14.133 and $22.613 in USD. Audio input consumes 7 tokens per second, and audio output consumes 12.5 tokens per second. One hour of speech in and speech out costs about $1.54 in Singapore before text and image tokens. The context window is 53,248 tokens, with 49,152 for input and 4,096 for output. Default rate limits are 10 requests and 100,000 tokens per minute. Model Studio lists function calling, structured outputs, batch inference and fine-tuning as unsupported.