Alibaba has introduced Qwen3.8-Omni-Flash, a multimodal AI model capable of processing text, images, audio, and video within a single architecture. The system supports a 1-million-token context window and is designed for tasks such as real-time dialogue, video editing, and audiovisual summarization.
The new model offers significant cost reductions for API access compared to the Qwen3.5-Omni-Plus version, with audio processing costs dropping by over 98% and combined audio-video processing by over 93%. Performance across 29 benchmarks has improved by more than 25% on average.
Qwen3.8-Omni-Flash features an agentic understanding mechanism that selectively analyzes media files based on user queries rather than sequential processing. This approach improved performance in OmniVideoBench and reduced token consumption per request by 45.7%.
In a test, the model successfully improved a smaller AI system, Qwen2.5-Omni-3B, by generating 3,413 training examples over 12 hours. This process reduced the character error rate for recognizing the Sichuan dialect from 25.79% to 15.30%.
The model supports 74 languages for speech recognition and 29 for generation. Additionally, Alibaba released the open-source Qwen-Live Harness environment for building interactive applications with low latency, utilizing WebSocket and WebRTC protocols.
