Alibaba’s Qwen team has released Qwen3.5-Omni, an omnimodal model that accepts text, images, audio and video as input and responds in text or speech in real time. The company positions it as a single unified system rather than a chain of separate models for speech recognition, vision and language.
The release comes in three sizes, Plus, Flash and Light, intended to cover different trade-offs between capability, speed and cost. Alibaba says the models support a context window of up to 256,000 tokens and were trained on tens of millions of hours of audio and video data alongside text.
Built for spoken conversation
Several of the new features target voice interaction. According to Alibaba, the model includes a mechanism that distinguishes between a user who is actually addressing it and background speech or incidental sounds, so that it does not respond to every noise it hears. The company says this makes spoken exchanges feel more natural and reduces interruptions.
A second feature keeps the text and speech output aligned, which Alibaba says reduces errors when the model reads out numbers, names or technical terms. The model also supports voice cloning: a user can provide a short sample of their own voice and have the model answer in that voice. Alibaba reports improved performance across 36 spoken languages and says the model outperforms several competing systems on audio understanding benchmarks, a claim based on the company’s own evaluations.
Code from a screen recording
One of the more unusual capabilities Alibaba highlighted is the ability to generate code from a video or screen recording. In the company’s demonstrations, the model watched a recording of an interface or workflow and produced code to reproduce it. Alibaba also says the model can access real-time information, allowing it to answer questions about current news or market data.
In the company’s examples, the model analyzed video clips quickly and returned structured summaries of their content. Independent evaluations of these capabilities were not available at launch.
Part of a broader push
Qwen3.5-Omni follows the Qwen3.5 family of language models and the earlier Qwen2.5-Omni and Qwen3-Omni releases, which established Alibaba’s approach of training a single model to handle multiple modalities natively. Alibaba has made many of its Qwen models available under open licenses, which has helped them gain adoption among developers outside China.
The launch places Alibaba in direct competition with Google’s Gemini and OpenAI’s GPT-4o-class models, which also handle speech, vision and text in a single system. The main differentiators Alibaba is pushing are multilingual speech quality, voice customization and the video-to-code capability, along with the option to run smaller versions at lower cost.
