Alibaba’s Qwen team has launched Qwen3.8-Omni-Flash, positioning it as their first omni-modal model built specifically for agentic workflows. Unlike previous iterations that processed media linearly, this model accepts text, images, audio, and video but outputs only text. The architecture relies on a “coarse-to-fine” reasoning strategy. Instead of digesting an entire two-hour video file, the agent analyzes the user’s question, identifies relevant segments, and gathers evidence iteratively. This targeted approach drops token consumption from 145,736 to 79,117 on the OmniVideoBench test, a 45.7% reduction. Accuracy simultaneously climbed from 63.4 to 67.8.
The model runs on the Qwen3.8-Flash-Next base, which originally shipped with open weights in August 2026. However, Qwen3.8-Omni-Flash is not open source. It is available immediately via hosted API on QwenCloud, Alibaba Cloud Model Studio, and Qwen Studio. Developers can access it through DashScope or OpenAI-compatible protocols, supporting function calling, web search, and structured outputs. The context window stretches to 1 million tokens, with a maximum input of 991K and output of 131K. Reasoning effort defaults to “xhigh,” though users can disable thinking entirely if speed outweighs depth.
Pricing reflects the efficiency gains. Input tokens cost $0.15 per million, while output tokens run $0.47 per million. Implicit cache hits drop to just $0.016 per million tokens. Alibaba claims these architectural changes slash audio input costs by over 98% per hour compared to Qwen3.5-Omni-Plus. Video input costs fall by roughly 89%. The model supports video files up to two hours and 2 GB, along with audio inputs spanning 113 languages. It handles stable results at up to 15 fps and supports multi-channel spatial audio.
To bridge the gap between text-only output and multimodal input, Alibaba open-sourced Qwen-MM-Plugins under Apache-2.0. This toolkit lets existing agent harnesses like Claude Code and Gemini CLI integrate multimodal skills such as omni-memory and omni-video2note. While the main model cannot yet natively ingest audio through local harnesses, the API routing handles that load. Independent benchmarks remain scarce, but Qwen reports average score improvements of over 25% across 29 evaluations compared to its predecessor. Whether these internal metrics hold up against Google’s Gemini 3.8 Flash in real-world deployment remains to be seen.
