Alibaba releases Qwen2.5-Omni multimodal model

Alibaba’s Qwen team released Qwen2.5-Omni, a 7-billion-parameter open-weight model designed to handle text, images, audio and video as input and to generate both text and streaming natural speech as output — a single end-to-end system rather than separate models chained together for each modality. Weights were published under the Apache 2.0 licence on Hugging Face, ModelScope and GitHub.

The model used what Alibaba called a “Thinker-Talker” architecture: a “Thinker” component processed multimodal input and produced text, while a separate “Talker” component converted that output into streaming speech, with a position-embedding scheme the team named TMRoPE used to keep video and audio timestamps synchronised. Alibaba reported that the model matched or exceeded similarly sized single-modality specialists — outperforming its own Qwen2-Audio on audio tasks and matching Qwen2.5-VL-7B on vision benchmarks — and reported state-of-the-art results among open models on OmniBench, a benchmark for combined multimodal understanding.

The release extended a run of open-weight models from Chinese labs — Alibaba’s own Qwen series, DeepSeek, and others — competing directly on multimodality and real-time voice interaction, capabilities that had mostly been the preserve of closed, subscription-gated products such as OpenAI’s Advanced Voice Mode. A small, freely licensed model claiming comparable real-time voice performance lowered the barrier for developers to build voice-native applications without a closed API dependency.