Qwen's next-generation native omnimodal model, Qwen 3.8-Omni-Flash, officially launched on September 18. The model supports text, image, audio and video input, handles contexts of up to 1M tokens, and can autonomously plan tasks, call tools and complete creation and delivery based on its understanding of content.
Qwen 3.8-Omni-Flash enhances audio and video processing capabilities
In terms of functionality, Qwen 3.8-Omni-Flash retains its capabilities in coding, text processing and GUI operations while further expanding its ability to handle audio and video tasks. For long-form audio and video understanding, the model can handle hours of video content, determine which visuals and audio segments to focus on based on the question, and pinpoint key information through multiple rounds of investigation. Combined with tool use, the model can also extract meeting minutes, action items and risk information from meetings, then carry out follow-up tasks such as sending emails, writing code or retrieving multimodal data to generate reports.
On the audio and video creation side, the model can be applied to scenarios such as music video production, short-drama translation and long-film commentary. For example, in music video production, the model can understand song structure, rhythm and emotion, and output timestamped line-level lyrics; in short-drama translation, it can perform character identification, colloquial translation, voice cloning, audio track remixing and final quality check; in long-film processing, it can carry out plot distillation, commentary planning, voice-over and scoring, editing and rendering, and final quality check.
In addition, the model can transform long-form video content into reusable information assets, including automatically organizing knowledge points, breaking down steps, generating illustrated PDF notes, and further converting them into validated and evaluated Agent Skills. The real-time interaction version, Qwen 3.8-Omni-Flash-Realtime, can perform perception and response simultaneously as audio and video streams are fed in, making it suitable for scenarios such as spoken-language practice, spatial audio perception and intelligent customer service.
On the supporting side, the company has expanded Qwen-MM-Plugins in parallel, adding multiple capabilities for audio and video understanding and creation, and open-sourced Qwen-Live Harness for sustained, real-time audio and video interaction and task execution. According to official data, across 30 cumulative benchmarks, Qwen 3.8-Omni-Flash scored an average of more than 26% higher than its predecessor Qwen 3.5-Omni-Plus. Meanwhile, the API price per hour for audio input has dropped by more than 98%, and the price per hour for audio and video input has dropped by more than 93%.
The model is now available on the Qwen AI platform.