Alibaba Qwen Releases Qwen3.8-Omni-Flash: A 1M-Context Omni-Modal Model Built Around Agentic Audio-Video Understanding and Tool Use
MarkTechPost Asif Razzaq
Alibaba’s Qwen launched Qwen3.8-Omni-Flash, an API-only model for text, images, audio, and video. It uses tools, 1M context, and no open weights at launch.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Alibaba’s Qwen team has put out Qwen3.8-Omni-Flash, a new omni-modal model that takes text, images, audio, and video and answers in text. The pitch is not just that it can read and watch things. It’s that it can decide what matters, use tools, and finish the job in one flow.
That makes the deployment story pretty clear. The model is available today through hosted APIs on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio. No open weights were announced at launch, so anyone hoping to run it locally is out of luck for now. The API supports both DashScope and OpenAI-style calls, plus Chat Completions and the Responses API.
Under the hood, Qwen says the model sits on the Qwen3.8-Flash-Next architecture. The context window is 1M tokens, with QwenCloud listing 991K maximum input and 131K maximum output. Reasoning can go up to 262K tokens. Thinking is on by default, and developers can turn it off by setting reasoning_effort to none.
The more interesting claim is the agentic video path. Instead of chewing through an entire long clip from start to finish, the system starts with the question, chooses what to watch and hear, then gathers evidence in several coarse-to-fine passes. Qwen says that approach lifted OmniVideoBench accuracy from 63.4 to 67.8 while cutting token use from 145,736 to 79,117.
The benchmark story is entirely Qwen’s own, so it should be read that way. Still, the numbers are broad enough to matter. Across 29 evaluations, Qwen says the model averages more than 25% better than Qwen3.5-Omni-Plus. It also claims audio-visual performance close to Gemini 3.8 Flash, and overall audio performance above it.
Pricing is aggressive too: QwenCloud lists $0.15 per 1M input tokens and $0.47 per 1M output tokens. The docs also say the model can handle video files up to 2 hours and 2 GB by URL, audio up to 3 hours, and audio input across 113 languages and dialects. On top of that, Qwen is open-sourcing Qwen-MM-Plugins under Apache-2.0 to help agent harnesses handle multimodal work, though the README says audio still gets routed through the API for now.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI that actually looks useful: fewer magic demos, more systems that can watch, listen, call tools, and get out of the way. The catch is the usual one — the shiny multimodal part arrives as a hosted service first, while the open bits arrive later, if they arrive at all. Convenient for Alibaba, less romantic for everyone else.
Read more about this at: MarkTechPost