Qwen ships Qwen3.8-Omni-Flash with 1M-token context and agentic video perception
- Alibaba's Qwen team released Qwen3.8-Omni-Flash, a native omnimodal model handling text, images, audio and video in one workflow, with input near 991,000 tokens and single requests up to 2 hours of video or 3 hours of audio.
- Its sparse mixture-of-experts design holds roughly 125B total parameters with about 6B active per token across 512 experts, built directly on the Qwen3.8-Flash-Next architecture rather than pairing a text model with separate perception encoders.
- Qwen's agentic perception picks which parts of a recording to inspect in coarse-to-fine passes, lifting internal long-video benchmark accuracy from 63.4 to 67.8 while cutting tokens per query from about 145,700 to 79,100, a drop of roughly 46%.
- API pricing is $0.15 per million input tokens and $0.47 per million output tokens (cached input $0.016), and per hour of media Qwen reports audio-input costs down about 98% and combined audio-video costs down roughly 89% versus Qwen3.5-Omni-Plus.
- The model ships as a hosted service on Qwen Chat, Alibaba Cloud Model Studio and QwenCloud API across six regions with no open weights, while the companion Qwen-MM-Plugins toolkit went out under Apache 2.0 to let Claude Code, Codex and Gemini CLI read images, video frames and audio locally.