Alibaba releases Qwen2.5-Omni, a 7B model that sees, hears and talks

Background

Alibaba's audio models had been developed in steps before Qwen2.5-Omni. Qwen-Audio was pre-trained on more than 30 tasks covering speech, natural sounds, music and songs, and Qwen2-Audio came later. Qwen2.5-Omni brought speech, images and video into one model.^[4][2]

What happened

On 26 March 2025 Alibaba released Qwen2.5-Omni-7B under the Apache 2.0 license. A 3B version was released under the Qwen Research license. The technical report was posted to arXiv the same day.^[2][3]

The model is end-to-end and multimodal. It perceives text, images, audio and video, and it responds with text and natural speech.^[1]

Key details

It uses a Thinker-Talker architecture. The Thinker produces text, and the Talker turns the text into streaming speech.^[1]

Its position encoding, Time-aligned Multimodal RoPE, synchronises video timestamps with audio.^[1]

The model is designed for fully real-time interaction, with chunked input and immediate output. It offers two voices, Chelsie and Ethan.^[1]

On OmniBench the model averaged 56.13%, against 42.91% for Gemini-1.5-Pro.^[1] The team also reported that it matched Qwen2.5-VL-7B on vision tasks and outperformed Qwen2-Audio on audio tasks.^[1]

Some of the terms need explaining. End-to-end means that one model handles the whole task, rather than passing work between separate systems for listening, understanding and speaking. In a Thinker-Talker design, the first part decides what to say in text, and the second part turns that text into speech as a stream.^[1]

A multimodal model accepts more than one kind of input, here text, images, audio and video. OmniBench is a test that checks how well a model combines these inputs when it answers questions.^[1]

Real-time interaction means the model can respond while the user is still speaking or showing content. Chunked input means the model receives its input in pieces rather than all at once. The release also stressed immediate output.^[1]

Impact

The release showed one model handling speech input and speech output alongside text, images and video. In September 2025 Qwen2.5-Omni was succeeded by Qwen3-Omni.^[2]