Alibaba Showcases Qwen Model Advances, Next-Gen Video Model Due in November

At the 2026 Apsara Conference in Hangzhou on September 22, Alibaba detailed the latest progress in its Qwen family of large language models, part of a broader push spanning AI models, AI chips, and AI cloud.

The company said Qwen4, trained on a new architecture, is currently in development, with Qwen4.5 and Qwen5 also advancing and planned to eventually scale to 5-10 trillion parameters. Alibaba also showcased Qwen-Image-3.1 for image generation, aimed at commercial scenarios such as e-commerce and design, and released Happy Shrimp 1.1, a music generation model supporting over 100 music styles and more than 10 languages. In addition, the latest preview of its world model, HappyOyster 2.0, was introduced with improved real-time interaction, memory, and physical-rule following, while the omni-modal model Qwen3.8-Omni made its debut within a unified understanding framework for audio, video, and other modalities.

Alibaba also said its next-generation video generation model will be released in November, evolving toward longer, more controllable, and more complete outputs. The conference additionally saw the debut of T-Head's self-developed AI chip Zhenwu V900 and the company's target of operating more than 20GW of data centers globally by 2032.