Qwen3.8-Omni-Flash

Qwen3.8-Omni-Flash was released by Alibaba on 18 September 2026 as a natively omni-modal model: a single request can carry text, images, audio and video together, and it answers in text. The context window is one million tokens, with a maximum input of about 991,000 tokens and up to 131,000 tokens of output, which is enough to hold roughly an hour of continuous audio, or audio and video, inside one call. It recognises speech in 74 languages. Pricing is $0.15 per million input tokens, $0.016 per million cached input tokens and $0.47 per million output tokens; Alibaba frames the change as a cost cut rather than a capability jump, putting an hour of audio input at 98% less than its predecessor and an hour of combined audio and video at 93% less, and the domestic Bailian rate for multimodal input at ¥0.8 per million against roughly ¥18 before. It is available through the API only at launch — QwenCloud, Alibaba Cloud Model Studio and Qwen Studio, under the model id qwen3-8-omni-flash — across six regions including Singapore, Frankfurt and Virginia. The weights are not released. Alibaba reports accuracy on OmniVideoBench rising from 63.4 to 67.8 while using about 45.7% fewer tokens; that is the vendor's own figure, the 63.4 belongs to the earlier model, and no other model on this board has been measured on that benchmark, so it is recorded here in words and not stored as a score. Every figure in this description is Alibaba's own, taken from two independent reports published on 18 September that cite the company's launch materials and agree with each other; Alibaba's announcement page could not be read directly, as it renders only in a browser.