Alibaba publishes a model that thinks natively about audio and video and knows how to call tools. Unlike previous Qwen, it remains reserved for the API.
On September 18, 2026, Alibaba released Qwen3.8-Omni-Flash, an omnimodal model designed for agentic uses. It accepts a context window of one million tokens and processes sound and moving image without going through a prior transcription step: the reasoning applies directly to the audio-video stream.
The team presents this model as a cursor shift: it is no longer just a question of understanding multimodal content, but of planning a series of tasks, calling external tools and carrying out production work. The model can be accessed on QwenCloud, Alibaba Cloud Model Studio and Qwen Studio.
A notable difference with the usual trajectory of the Qwen family: no major game was released at launch. Self-hosting is therefore excluded for the moment, which contrasts with the opening strategy that had made the reputation of the series.
Read the full article on the original source
Go to the source →