Alibaba says its Qwen family has passed 470 open models and 4 billion downloads

What Tsai said at Wuzhen

On November 3, the 2026 World Internet Conference Wuzhen Summit opened in Wuzhen, in China's Zhejiang Province. Alibaba co-founder and chairman Joe Tsai spoke at the opening ceremony and gave the latest scale of Alibaba's open model family, Qwen. As BigGo Finance reported, Qwen has open-sourced more than 470 models, with cumulative global downloads above 4 billion and more than 330,000 derivative models.

"Open-sourced" here means the models are published so others can download and use them. A derivative model is one that outside developers have built by adapting an existing Qwen model for their own purposes.

Tsai also argued that Chinese tech companies have stepped up their open-sourcing in recent years. In his telling, they now lead the world on key measures such as downloads and derivative model counts. He described Chinese open models as strong performers that run efficiently, and said they are well regarded globally.

How to read the numbers

Two comparisons help. If 470 models had been released one per day, that would take more than a year. And 4 billion downloads is roughly one for every two people alive, using about eight billion as the world population. Both are our own back-of-envelope estimates, not Tsai's.

The caveats matter more. A download counts a file being fetched, not a person and not a business running the model in production. We have also seen only one outlet relaying a speech, not a public registry count, so the figures are reported rather than independently confirmed.

A total of 470 almost certainly includes many sizes and variants of the same underlying models. That is our analysis, not something the source states. Licenses can also differ from one model to the next, so the model card, the short documentation page that comes with each release, is worth reading before you build on one.

Tsai's case for open source

Tsai framed the speech around a supply problem. Demand for AI keeps rising as it spreads into manufacturing, healthcare, public services, office work and scientific research, he said. Computing power, the hardware that runs these systems, falls short of that demand, and he expects the gap to be hard to close soon. His line was that "only open source can drive widespread adoption", and he tied that adoption to lasting, inclusive progress.

He made the case at three levels. For companies, open models can be run in their own data centers, which helps protect their data. For the wider industry, open models lower research and development costs for small and medium-sized firms, which he said matters especially in developing countries.

For countries and regions, he argued that open source keeps AI from being the preserve of a few tech powers and helps narrow the gap in access to intelligence between places.

The business question behind the milestone

Analysts quoted in the report say open models are changing how AI companies compete. Raw model capability is no longer the only defense. A developer community, low running costs and control over data are becoming central to how businesses choose a model.

For a cloud provider, the logic runs like a loop. Downloads and derivative models create demand for inference, which means the computing needed to run a model and get answers, and for enterprise services. For small firms and developers, a free model they can run themselves lowers the cost of building AI products.

The report is plain about what remains unsettled. Once a model is given away, a company has to earn money from cloud services, technical support and custom delivery. Whether Qwen's scale advantage turns into steady commercial returns is, in the report's words, the focus of market attention to come.