Qwen3.8-Omni-Flash launches: watches videos on its own, calls tools, and delivers finished cuts

动察 Beating AI News Flash: Alibaba's Qwen launches natively omnimodal model Qwen3.8-Omni-Flash, supporting text, images, audio, video, and 1M context.

This generation has clearly strengthened audio-video Agent capabilities. The model can find information from videos on its own, plan tasks, and then call external tools to complete editing, dubbing, and rendering. The official showcase includes scenarios such as long meeting processing, short drama translation, movie commentary, MV production, and video-to-illustrated-note conversion.

When faced with several-hour-long videos, it does not need to process everything from beginning to end. The model first determines where the answer may be, then progressively locates key segments. On OmniVideoBench, accuracy rose from 63.4 to 67.8, while Token consumption dropped from 145,736 to 79,117, a decrease of 45.7%.

Qwen says that Qwen3.8-Omni-Flash's average score across 30 evaluations improved by more than 26% over Qwen3.5-Omni-Plus, with audio-video capabilities approaching Gemini 3.8 Flash and audio capabilities overall surpassing the latter. Audio input costs fell by more than 98%, and audio-video input fell by more than 93%.

Currently, the Qwen3.8-Omni-Flash model weights have not yet been made public, and only the API is open. Qwen has also open-sourced Qwen-MM-Plugins, which can be installed into Agents such as Claude Code, Codex, Gemini CLI, and Qwen Code, equipping them with image, audio, long-video processing, and video editing capabilities.

Another set, Qwen-Live Harness, is more like a real-time scheduling layer. Users can directly speak with the real-time omnimodal model, then hand complex tasks to background Agents such as Gemini CLI, Claude Code, and Codex for continuous execution. It will track task progress and proactively come back to tell you the results when done.