Qwen3-Omni

Qwen3-Omni is Alibaba's natively end-to-end omni-modal model, released in September 2025. It accepts text, images, audio, and video as input and generates both text and speech output, aimed at real-time voice assistants and multimodal agents. The architecture uses a Thinker-Talker design: a Thinker module handles cross-modal reasoning while a Talker module generates natural speech, keeping audio-only latency around 211ms per Alibaba's report. It is a 30B-parameter mixture-of-experts model with about 3B active parameters per token. Text support covers 119 languages, with 19 languages for speech input and 10 for speech output, including regional dialects such as Cantonese. Three variants shipped: Instruct, Thinking, and Captioner. Alibaba reports state-of-the-art results on 22 of 36 audio/video benchmarks, a 1.7% word error rate on speech recognition, and 90.2% on instruction following; treat these as team-reported until independently reproduced. Weights are on Hugging Face and ModelScope under Apache 2.0, and a demo runs on chat.qwen.ai. As with all omni-modal models, real-time speech quality depends heavily on network conditions and the deployment stack, not just the checkpoint.

Team-reported: SOTA on 22/36 audio-video benchmarks, 1.7% WER speech recognition, 58.7% AIME 2025, 90.2% instruction following