Alibaba's Qwen team has released Qwen3.8-LiveTranslate, a hosted simultaneous interpretation model that reduces average length-adaptive lag (LAAL) to 2.3 seconds across 60 input languages and 29 spoken-output languages. The model is available exclusively through Alibaba Cloud Model Studio's WebSocket API under the model ID qwen3.8-livetranslate-flash-realtime.
The update introduces three core features: real-time speaker separation with voice cloning that preserves each speaker's timbre, synchronized bilingual output that streams source transcripts alongside translations, and long-context disambiguation that uses conversation history to keep names and terminology consistent across sessions.
| Benchmark | Qwen3.8-LiveTranslate | Qwen3.5-LiveTranslate | Seed LiveInterpret 2.0 | GPT-Realtime-Translate | Gemini 3.5 Live translate |
|---|---|---|---|---|---|
| xCOMET-XXL | 85.7 | 83.0 | 79.4 | 73.5 | 65.2 |
| LAAL | 2.3 | 2.5 | 2.8 | 2.8 | 3.1 |
| ASR WER | 6.6 | 7.0 | 11.4 | — | — |
| UTMOS | 4.0 | 3.9 | 3.4 | 3.4 | 3.0 |
| — = not reported by vendor in this release |
Figures are vendor-disclosed; higher is better unless noted. Independent replication pending.
Confirmed
- Architecture: Hybrid-MoE Thinker–Talker design with an Interleave architecture that processes video, audio, source text, and translation in a single causal sequence; the Thinker handles streaming understanding and text output, while the Talker synthesizes speech preserving the original speaker's voice.
- Latency: LAAL of 2.3 seconds, down from 2.8 seconds in the previous Qwen3.5 release, measured on the Omnilingua-MSpeaker benchmark (14 language directions) and FLEURS audio test set (70 language directions).
- Language coverage: 60 input languages understood, 29 languages for speech output.
- Context window: 53,248 tokens (max input 49,152, max output 4,096).
- Pricing (international/Singapore region): Audio input $7.50 per million tokens, image input $0.55 per million tokens, text output $20 per million tokens, audio output $30 per million tokens. China (Beijing) region: audio input $5.653, image input $0.466, text output $14.133, audio output $22.613 per million tokens.
- Rate limits: 10 requests per minute (RPM), 100,000 tokens per minute (TPM) in both regions.
- Audio tokenization: 12.5 tokens per second for both input and output (750 tokens per minute). One minute of continuous audio input plus one minute of translated audio output costs approximately $0.0281 before text and video charges.
- Integration: Single WebSocket endpoint with session configuration for target language, output modalities (text, audio, or both), input audio transcription (optional, via qwen3-asr-flash-realtime), and hotword corpus for domain-specific terminology.
- Model access: Closed weights, API-only; no fine-tuning, function calling, structured outputs, web search, prefix completion, context caching, or batch inference supported.
Unknown
- Independent replication of the 2.3-second LAAL claim and the reported >94% offline translation quality retention in real-time conditions.
- End-to-end application latency once capture, network transit, synthesis, buffering, and playback are included (LAAL measures model-internal delay only).
- Diarization Error Rate (DER) improvements in absolute terms; the blog states Qwen3.8 outperforms mainstream systems on DER but does not publish a numeric value.
- Availability timeline for fine-tuning, function calling, or structured output support.
- Whether the model will be released with open weights or remain API-only.
Our take
Qwen3.8-LiveTranslate collapses what used to require separate ASR–MT–TTS components into a single WebSocket call, lowering integration friction for conferencing, livestreaming, and classroom tools. The half-second latency gain is real, but the practical edge is speaker-aware voice cloning and synchronized bilingual output—features that cut the engineering burden of multi-speaker translation UIs. At roughly $1.69 per hour of bidirectional audio, pricing is competitive for hosted real-time translation, though token-based billing means costs climb with silence unless applications filter voice activity upstream.
