Sources
Qwen 3.5 Omni: Alibaba Ships Natively Omnimodal Model — 100M Hours Training, 113 Languages, Voice Cloning, Semantic Interruption
Released March 30, Qwen3.5-Omni processes text, images, audio, and video simultaneously in a single pipeline — no modality conversion needed. Trained on 100M+ hours of audio-visual data across three sizes (Plus, Flash, Light) with 256K context, it supports 113 speech-recognition languages and 36 spoken languages, achieves SOTA on 215 audio/AV understanding subtasks, and introduces semantic interruption (distinguishing 'uh-huh' from genuine interjections) plus voice cloning from a single sample. This puts direct competitive pressure on ElevenLabs and Gemini 3.1 Pro's audio capabilities simultaneously.
↳ Follow the thread