Qwen2.5-Omni-7B

Qwen2.5-Omni-7B is Alibaba's end-to-end omni-modal model, released March 2025. It accepts text, images, audio, and video and produces real-time text and natural speech, using the Thinker-Talker architecture: a Thinker transformer fuses the modalities and generates text, while a Talker streams speech from the Thinker's hidden states. Time-aligned multimodal RoPE keeps video and audio in sync, and block-wise streaming enables low-latency spoken replies.

Alibaba released it under Apache 2.0 and positioned it as a foundation for cost-effective AI agents, including assistive tech such as real-time audio descriptions for visually impaired users. At 7B parameters it runs on capable edge devices, though Alibaba's own figures show video inference is memory-hungry (tens of GB of VRAM for minute-long clips).