Qwen: Qwen3.8 Omni Flash

Qwen3.8-Omni-Flash is Alibaba Qwen's native omnimodal agent model for real-world productivity. It unifies text, image, audio, and video in a single Thinker-Talker sparse MoE stack built on the Qwen3.8-Next language backbone, with upgraded AuT and Spatial AuT encoders for general and spatial audio. The model supports up to one million tokens of context for long-form audiovisual reasoning, meeting understanding, and multi-step agent planning while keeping text performance in line with Qwen3.8-Flash-class models.

Training combines native multimodal co-training, specialist distillation, and reinforcement learning on long-horizon agent data so reasoning and tool-use skills transfer from text to vision and audio-visual tasks. Evaluations in the Qwen3.8-Omni technical report highlight strong coding and agent scores (for example SWE-Bench-Pro 63.3, LiveCodeBench v6 92.6, GPQA Diamond 91.0, HLE 36.5) alongside large gains over Qwen3.5-Omni-Plus on audio, video, and multimodal agent benchmarks.

The model is offered primarily through DashScope and Qwen Cloud APIs rather than as a fully open weight release. Alibaba pairs it with Qwen-MM-Plugins and Qwen-Live-Harness so developers can plug audiovisual perception, selective retrieval, and realtime interaction into existing agent frameworks for workflows such as video editing, localization, film commentary, and omnimodal research.