Qwen launches Omni-Flash: first omni-modal model with built-in agentic capabilities

Alibaba's Qwen group has released Qwen3.8-Omni-Flash, the company's first omni-modal model designed from the ground up around agentic capabilities. Unlike models that merely "understand" video and audio, the new model is meant to link comprehension, task planning and tool invocation in a single flow: analyse what is seen and heard, plan the next steps, run external tools and return a finished result. The company says this enables scenarios such as automated vlog editing, short-form video translation or feature-film summarisation without breaking the process across multiple separate models.

The headline figure in the announcement is a direct comparison with Google's Gemini 3.8 Flash: Qwen claims the new model "approaches" the rival's audio-video capabilities and logs an average 19.5-point improvement in agent performance on two dedicated benchmarks, WildClawBench-MM and UniClawBench. Both tests measure the ability to orchestrate tools across long, multi-step workflows, not just answer single questions. Full benchmark results and testing conditions have not yet been published, so the numbers come from the vendor alone.

The model ships with a 1-million-token context window, and Qwen highlights an "agentic perception" capability that lets it scan long videos and locate key moments with higher accuracy. According to internal testing on OmniVideoBench, this approach consumes 51.8% fewer tokens than static understanding of the same content — a meaningful saving when dealing with long-form video input. The practical upshot: more raw material fits into the same context budget, or costs can be lowered.

On the economics side, Qwen reports a roughly 89% reduction in video-input cost versus Qwen3.5-Omni-Plus, the previous generation in the family. That makes long-form omni-modal workflows far more accessible for developers building applications on top of the model. The new pricing, combined with the expanded context window, targets use cases such as lecture analysis, recorded meetings or enterprise video libraries where input volume has until now been a budget bottleneck.

To ease application development, Qwen is open-sourcing two companion components: Qwen-MM-Plugins, a plug-in library for multi-modal media handling, and Qwen-Live Harness, a framework for running and evaluating live workflows with the model. The first is already on GitHub; the second is expected shortly. The move signals an intention to turn the Omni ecosystem into a platform rather than a single model.

The model is available now through three channels: direct chat in Qwen Studio (chat.qwen.ai), a managed API on Alibaba Cloud (alibabacloud.com), and self-hosted deployment via Qwencloud. Technical documentation and the official blog detail rate limits and the request structure for combined audio-video input.