Qwen3.8-Omni-Flash takes video and audio in a 1M window

Alibaba’s new omni model, Qwen3.8-Omni-Flash, takes text, images, audio and video as input and holds all of it in a 1M-token context window. The company launched it on September 18 and said in a post on Threads that it performs close to Gemini 3.8 Flash on WildClawBench-MM and UniClawBench.

Alibaba named the two benchmarks and the rival model but gave no scores for either side. The model also supports function calling and tool use, so it can call external tools from inside the same session where it reads a clip or a voice recording.

Much of that spec sheet already belongs to Qwen3.8-Flash, the 125B Mixture-of-Experts model Alibaba released on August 26 and pitched as a preview of Qwen4’s architecture. The QwenCloud documentation lists that model with 6B parameters active per token, a native 1M-token context, function calling, and understanding of text, images and video. So the 1M window and the tool use are both carried over, and the one input type the Omni model adds over the hosted Flash is audio.

The Omni models are their own line at Qwen, and Qwen’s Hugging Face collections keep a Qwen3-Omni family apart from the Qwen3.8 releases. The line started with Qwen2.5-Omni on March 26, 2025. According to its technical report, that model split the job in two. A language model called Thinker writes the text, and a second model, Talker, turns Thinker’s hidden states into audio tokens, with both trained end to end so the model can answer in streaming speech as well as text.