Alibaba Qwen ships Qwen3.8-Omni-Flash: a 1M-context agentic audio-video model
On 18 September Alibaba's Qwen team unveiled Qwen3.8-Omni-Flash, which processes text, images, audio and video in a single request. It ships with a 1M-token context, tool calling and a thinking mode, and cuts audio-input cost by 98% versus its predecessor.
Alibaba's Qwen team unveiled Qwen3.8-Omni-Flash on 18 September. The company presents it as its first omni-modal model built around agentic capabilities: it takes text, images, audio and video in a single request and returns text. The goal is to fold four steps into one model, understand the content, plan the task, apply it with tools and deliver the result.
What it brings
The model ships with a 1M-token context window, and supports a thinking mode, tool calling and web search. Long-video analysis, meeting summaries, video research and multimodal tool use are the headline jobs. Alibaba says it can coordinate tools across long workflows to auto-edit vlogs, translate short videos and turn films into summaries.
Performance and cost
Alibaba says the new model raised its average score by more than 26% across 30 evaluations versus the earlier Qwen3.5-Omni-Plus; it gained an average of 19.5 points in audio-video agent performance across WildClawBench-MM and UniClawBench, and approaches Gemini 3.8 Flash on audio-video. On OmniVideoBench it locates key moments with higher accuracy while using 51.8% fewer tokens than static understanding. On cost, an hour of audio input is 98% cheaper than before, and audio and video combined 93% cheaper.
Access
The model is listed on Alibaba Cloud Model Studio at 0.8 RMB per million input tokens and 2.7 RMB per million output tokens, available in Beijing, Singapore, Hong Kong, Tokyo, Frankfurt and Virginia. Qwen is also open-sourcing Qwen-MM-Plugins and a Qwen-Live Harness to help developers build around the model.
Sources: Qwen (Alibaba), MarkTechPost.
