Alibaba Launches Qwen 3.5

Alibaba Group recently launched Qwen 3.5, a 397-billion-parameter open-weight AI model with visual agentic features that enable it to execute actions autonomously across mobile and desktop platforms. 

The new model activates only 17 billion parameters during each forward pass through a hybrid mixture-of-experts architecture, delivering what Alibaba claims is 60 percent lower operating costs and eight times better efficiency for large-scale workloads compared to the previous generation. 

The company said Qwen 3.5 outperforms GPT-5.2, Claude 4 Opus and Gemini 3 Pro on a range of benchmarks, though these comparisons are self-reported and have not been independently verified.

Qwen 3.5 supports native multimodal inputs, processing up to two-hour videos and more than 200 languages and dialects, an increase from 82 in the prior version. 

Alibaba positioned the expansion as part of its global ambitions, with analysts noting that broad language support signals the company's intent to compete outside China. The model is accessible via Alibaba's Model Studio as a hosted version and is also available as open weights for download, customisation and deployment on user infrastructure.

The launch comes just days after ByteDance released Doubao 2.0, an upgraded chatbot that has attracted 200 million users in China and now dominates the domestic market. Alibaba's emphasis on agentic capabilities directly tracks the industry's shift from conversational assistants to autonomous systems that can carry out complex multi-step tasks without human oversight. 

The Qwen series has already surpassed 20 million downloads, and open-weight availability is intended to build a developer ecosystem and solidify loyalty. Alibaba has begun integrating agentic functionality into its own platforms, including Taobao and Cainiao, for automated shopping and logistics processes.