Qwen 3.8-Omni-Flash Launches: Independently Watches Videos, Adjusts Tools, and Delivers Finished Works

Beating AI Express News Brief: Alibaba’s Qwen launches native multimodal model Qwen3.8-Omni-Flash, supporting text, images, audio, video, and a 1M context window. This generation significantly enhances audio and video agent capabilities. The model can independently extract information from videos, plan tasks, and call external tools to complete editing, dubbing, and rendering. The official demonstrated use cases including long meeting processing, short drama translation, movie commentary, MV production, and video-to-text-and-image note conversion. For hours-long videos, it does not need to process the entire content from start to finish; instead, it first identifies where the relevant answer may lie, then gradually pinpoints key segments. On OmniVideoBench, accuracy has risen from 63.4 to 67.8, while token consumption dropped from 145,736 to 79,117, a 45.7% reduction. Qwen stated that Qwen3.8-Omni-Flash’s average score across 30 benchmarks is over 26% higher than Qwen3.5-Omni-Plus. Its audio and video capabilities are close to Gemini 3.8 Flash, and its overall audio performance surpasses the latter. Audio input costs have fallen by more than 98%, and audio-video input costs by over 93%. Currently, the model weights of Qwen3.8-Omni-Flash have not been released; only its API is available. Qwen also open-sourced Qwen-MM-Plugins, which can be integrated into agents like Claude Code, Codex, Gemini CLI, and Qwen Code, adding image, audio, long video processing, and video editing capabilities to these tools. Another component, Qwen-Live Harness, acts more like a real-time scheduling layer. Users can directly interact with the real-time multimodal model, then assign complex tasks to backend agents such as Gemini CLI, Claude Code, and Codex for continuous execution. It tracks task progress and proactively informs users of results once completed.