Alibaba's Qwen Releases Qwen3.8-LiveTranslate Simultaneous Interpretation Model, Outputting Source and Translated Text in the Same Frame

Summary

Alibaba's Qwen released the simultaneous interpretation model Qwen3.8-LiveTranslate on September 19. Built on an Interleave architecture with a Hybrid MoE Thinker–Talker dual-module design, it cuts average latency per character (LAAL) from 2.8 seconds to 2.3 seconds and adds real-time speaker diarization, same-frame bilingual output, and long-context disambiguation. The company says it supports 60 languages and leads the previous generation and mainstream real-time interpretation systems on translation quality, latency, speech recognition and speech synthesis across the Omnilingua-MSpeaker and FLEURS benchmarks.