Alibaba launches Qwen3.5-Omni for realtime multimodal AI

Alibaba’s Qwen team released Qwen3.5-Omni, a multimodal model that can take and generate text, images, audio, and video, and the company said it is available via offline API and real-time API.

The model supports a 256K context window, and can process long-form audio and video inputs.

Alibaba said the model was trained on multimodal data, including over 100 million hours of audio and video, and that it improved multilingual speech with recognition in 113 languages and dialects and speech generation in 36 languages.

Alibaba also claimed new real-time features such as semantic interruption, voice cloning, and voice control, and benchmark results put its audiovisual performance on par with Google Gemini 3.1 Pro.

🔗 Source: Pandaily

🧠 Food for thought

Implications, context, and why it matters.

This new model is the latest move in a long-term, full-stack AI strategy

Qwen3.5-Omni lands as part of a multi-year plan that includes a proposed RMB 380 billion investment in AI and cloud infrastructure over three years 1.
Alibaba has already open-sourced more than 300 models based on its Qwen and Wan foundation models, reaching over 600 million downloads 1.
The company calls itself a “full-stack AI service provider” by tying its Qwen models to its PAI computing infrastructure (Alibaba Cloud’s Platform for AI tooling for training and deploying models) and the Model Studio application platform, so it can support the full AI lifecycle for enterprise customers 2.

Powerful open models create new opportunities and new burdens for businesses

Alibaba released models like Qwen3-Omni under an Apache 2.0 license (a widely used open-source license that allows broad commercial reuse), putting pressure on proprietary systems and giving enterprises another route to limit vendor lock-in 3.
The move speeds up adoption of multi-model stacks, where teams combine open and closed models to match specific workloads 3.
The trade-off includes more in-house work on MLOps (machine learning operations), fine-tuning, and governance, since enterprise AI, security, and compliance teams still carry the burden of security, privacy, and compliance risk management regardless of where a model comes from 4.