Alibaba's Qwen Releases Live Translation AI Model with Lower Latency

Alibaba's Qwen has released a new live translation AI model called Qwen3.8-LiveTranslate. The model uses an Interleave architecture to improve real-time translation, cutting average latency from 2.8 seconds to 2.3 seconds and supporting 60 languages.

The new model adds three key features: real-time speaker separation for clear attribution of each sentence, simultaneous display of original and translated text on the same screen, and long-context disambiguation for more accurate translation of names and terms. It uses a Hybrid MoE-based Thinker-Talker dual-module design that connects streaming understanding, text output and speech generation while preserving the original speaker's voice.

On the Omnilingua-MSpeaker multi-speaker long audio test set, the model outperforms mainstream real-time translation systems in translation fidelity, fluency, conciseness and speaker separation error rate. On the FLEURS test set across 70 language directions, it also leads in translation quality, average latency, speech recognition accuracy and speech synthesis quality.