Qwen3-Omni-30B-A3B-Instruct

Qwen3-Omni-30B-A3B-Instruct was released by Alibaba on September 22, 2025, as the open-weight flagship of the Qwen3-Omni family. The 30B-parameter mixture-of-experts model activates about 10% of its parameters per inference and uses a Thinker-Talker architecture for ultra-low-latency multimodal interaction.

It accepts text, image, audio, and video input with support for 119 text languages and 19 speech input languages, and is served through standard vLLM-style stacks. Released under the Apache 2.0 license, Alibaba reports it outperforms closed-source rivals on a majority of its audio/video evaluation tasks, which should be treated as vendor-reported.