Alibaba's Qwen Launches Omni-Modal Model Qwen3.8-Omni-Flash, Rivaling Gemini 3.8 Flash in Audio-Video Capabilities

Alibaba's Qwen platform has officially launched its new native omni-modal model, Qwen3.8-Omni-Flash. It can handle text, images, audio, and video inputs at the same time, supports 1M context, and is now available on the Qwen AI platform. The company says it keeps the same text capabilities as models of its size while greatly improving omni-modal abilities over the previous generation.

Across 30 benchmarks, Qwen3.8-Omni-Flash improved its average score by over 26% compared to the previous generation. In audio-video agent, coding, and long-horizon tasks, WildClawBench-MM rose by 36.5 points and AgenticVBench by 22.3 points. Its audio-video ability is close to Gemini 3.8 Flash, while its audio ability overall surpasses it. API pricing for audio input dropped by over 98% per hour.

The model supports up to one hour of audio-video input, enabling speaker separation, transcription, meeting summary generation, and task analysis. With tool calling, it can also send emails or write code. The real-time version, Qwen3.8-Omni-Flash-Realtime, is the first omni-modal large model to support sound source localization, identifying direction and distance.