What shipped
Alibaba’s Qwen team released Qwen3.8-Omni-Flash on Thursday, a model that takes text, images, audio and video as input and returns text, TechNode reported.
The context window is 1 million tokens, with a maximum input of 991K and output of 131K, according to MarkTechPost. Reasoning extends to 262K tokens and is on by default. The model is built on the Qwen3.8-Flash-Next foundation released in August, and it accepts video up to two hours or 2GB and audio up to three hours.
The positioning
This is not framed as a captioning model. Qwen is selling it as the engine for long audio-video agents: planning across a recording, calling tools, and producing a finished artefact. Two supporting pieces shipped alongside it — Qwen-MM-Plugins and a Qwen-Live Harness for long-running and real-time workflows.
Function calling, web search, structured outputs and context caching are all supported. Speech output is not: for generated audio, Qwen directs users to the earlier Qwen3.5-Omni.
The numbers, and who reported them
Qwen says the model improves its average score by more than 25% across 29 evaluations compared with Qwen3.5-Omni-Plus, with gains including 36.5 points on WildClawBench-MM and 9.6 on OmniVideoBench. The company also claims audio performance above Gemini 3.8 Flash and audio-visual performance close to it.
All of those are Alibaba’s own figures, published by Alibaba, with no independent evaluation attached. They describe the direction of travel within the Qwen family, which is the part a vendor can credibly measure. The comparison to a competitor’s model is the part that needs an outside test before it means anything.
Price and availability
QwenCloud pricing is listed at $0.15 per million input tokens and $0.47 per million output tokens, with access through QwenCloud, Alibaba Cloud Model Studio and Qwen Studio. TechNode reports input pricing of RMB 0.8 per million tokens.
No open weights were released at launch. That is the detail worth noting, because Qwen’s reputation in the last two years has been built substantially on publishing them. A flagship omni-modal model held back as an API is a different commercial posture from the one the family is known for.
What to watch
The first useful signal will be an independent audio and video evaluation — Artificial Analysis, or an academic group, running the same tasks on Qwen3.8-Omni-Flash and Gemini 3.8 Flash. The second is whether weights follow, as they did for Flash-Next in August.
