Alibaba unveils live translation model Qwen3.8-LiveTranslate with enhanced real-time performance

On September 19, 2026, Alibaba's Qwen unveiled its latest live translation model, Qwen3.8-LiveTranslate, which improves real-time simultaneous translation through an Interleave architecture. The model reduces the average latency (LAAL) from 2.8 seconds to 2.3 seconds, enhancing accuracy, fluency, and conciseness. It supports 60 languages and introduces three new capabilities to expand its application in real-world scenarios. The model uses a Hybrid MoE-based Thinker–Talker dual-module design, integrating streaming understanding, text output, and speech generation. In evaluations on the Omnilingua-MSpeaker dataset, Qwen3.8-LiveTranslate outperformed industry-standard real-time translation systems in translation fidelity, fluency, conciseness, and speaker identification error rates. On the FLEURS audio test set, it also led in translation quality, latency, speech recognition accuracy, and speech synthesis quality across 70 language directions.