Alibaba unveils Qwen4-Max: approximately 2.4 trillion backbone parameters, but only about 71 billion activated per token. The model is very large, but each inference call uses only a fraction of its parameters.

The question BABA now needs to answer is whether this design can improve efficiency in real-world business operations. The new model supports a 1-million-token context window and is designed for long-running AI agents and multimodal tasks. The figure showing a roughly 60% reduction in per-task costs comes from Alibaba’s own evaluations; it cannot be directly equated with customers’ actual costs or the company’s profits.

Before the U.S. market opens on November 19, Alibaba will report quarterly results for the period ended September 30. The consensus forecast is for revenue to grow about 8.9% year over year to 269.85 billion yuan. Next, investors will be watching whether growth in AI cloud services, investment, and cash flow add up: if the models require less computing power, will the gains show up on Alibaba’s income statement, or be passed on to customers through lower service prices?