Alibaba Releases Qwen3.8-LiveTranslate Cutting Lag to 2.3s

The Alibaba Qwen team has officially released Qwen3.8-LiveTranslate, a state-of-the-art real-time interpretation model built to address the classic tradeoff in simultaneous translation between waiting for complete context and minimizing listener delay. The model rebuilds this processing loop utilizing an innovative Interleave architecture.

Building on the Qwen-Omni stack, this new release integrates large-scale multimodal data, cross-language and cross-modal alignment, alongside visual enhancement features. This allows the system to process both audio inputs and optional visual cues—such as speaker gestures, lip movements, and on-screen text—to maintain high accuracy in noisy environments or when encountering ambiguous terminology, with documentation recommending no more than two images per second.

Stock photo for illustration only, not from the actual event

Regarding language coverage, the model comprehends 60 languages in total, returning both audio and text for 29 of them while outputting text-only for the remaining 31. Supported speech output includes Chinese, English, Arabic, German, French, Spanish, Japanese, Korean, and Hindi. Furthermore, development teams can configure up to 1,000 hotwords to map specific source terminology directly to fixed target translations.