A company whose machines are trained to predict the next word has left one for us to supply. On Friday (November 6), Alibaba’s Qwen released a poster without a model name or a benchmark. Birds gather above a dark sea, their flock taking the shape of a fish.
Beneath them, one classical Chinese sentence stops short of its final character. Three English words accompany it: The becoming begins. The teaser invites speculation about Qwen4, but announces neither a release date nor a technical specification.
北冥有鱼，其名为 _
In the northern sea there is a fish. Its name is …
The answer is 鲲, Kun, a name familiar to readers of Zhuangzi’s Free and Easy Wandering. Kun is so vast that no one knows how many thousands of li it spans. It transforms into Peng, a bird whose wings stretch like clouds across the sky. The fish does not simply grow larger. It leaves the element in which it has always lived.
There is an older irony in the name. Classical Chinese dictionaries associate kun with fish roe, the beginning of a fish; Zhuangzi gives the word to something too vast to measure. Whether he meant to make a joke is unknowable, but the effect survives translation.
What looks small may already belong to an entirely different scale. Qwen’s ordinary-sized models, running on machines far from Alibaba’s own cloud, have made a similar lesson commercially awkward for those who thought the frontier could be defined by the biggest laboratories alone.
Yet Zhuangzi interrupts the wonder with an unromantic observation. A mustard seed will float in a little water; a cup will not. Likewise, the great bird cannot rise until enough wind has accumulated beneath its wings. Its size makes its dependence greater. There is power in the transformation, but the wind must come from somewhere.
Forty-five days before the poster, at Alibaba’s September 22 Apsara Conference, chief executive Eddie Wu gave a remarkably concrete account of that wind. Machine intelligence, he said, should be understood as “an entirely different species,” no longer merely an imitation of human capabilities.
In his published address, he described chips as the power plants of an intelligence economy and cloud infrastructure as the grid distributing its output. He introduced the Zhenwu V900 accelerator, described clusters of up to 500,000 cards and set an ambition of more than 20 gigawatts of data-center capacity by 2032.
On the poster, the bird is beginning to fly. In September, the company was piling up air.
This is also a useful way to understand the contest over AI between the United States and China. Washington’s restrictions on advanced chips have constrained Chinese developers, while sharpening the incentive to build alternatives.
According to IDC figures reported by Reuters on October 13, Nvidia held 34% of China’s data-center accelerator shipments in the first half of 2026, against Huawei’s 28%. Nvidia still led, but its shipments were down 44% from a year earlier. Export policy has become part of the weather in which these companies work.
Hardware is only half the story. In an April study of open models, Nathan Lambert and Florian Brand found that Qwen had amassed about 942 million cumulative downloads by March 2026.
In February, 69% of newly uploaded model derivatives in their measured ecosystem were built on Qwen. The figures describe adoption, not a verdict on which model is smartest. They reveal how much development can happen after the original weights leave the laboratory.
And the reach is no longer confined to developers. Apple announced a China developer beta beginning November 4, with Alibaba’s Qwen involved in the China-specific Apple Intelligence offering. A model family that can be downloaded and modified is also finding a place inside some of the most tightly controlled consumer technology in the world.
Wu’s September speech made room for this apparent contradiction. He praised the existing open-weight Qwen-27B and promised better models for local deployment, while outlining a major expansion of Alibaba’s own cloud. The company can help people keep intelligence on their machines and still build the facilities that supply intelligence to millions of others.
To see why that matters, consider a small business prohibited from sending confidential documents to an overseas cloud service. For some tasks, a capable model operating on its own computers is a different proposition from a clever model delivered by subscription.
The same is true of researchers who need to inspect behavior and developers who work in languages too small a market to attract sustained commercial attention. Accessibility is not an abstract virtue when it changes who gets to build.
There is a practical distinction between renting intelligence and keeping it. An API account depends on a provider’s pricing, permissions and continued service. Weights stored locally can keep working without a connection to that provider.
They still require electricity, hardware and the judgment of whoever runs them. Nor does a public model release make the infrastructure needed to train the next frontier system any less expensive.
The fish in Qwen’s artwork offers one further twist on the old tale. Zhuangzi’s Kun is a single creature; the poster’s fish consists of many birds. Look closely, and the vast outline is made of small independent movements.
Qwen’s derivatives are created by fine-tuners, researchers and engineers whom Alibaba may never meet. Their work makes a foundation model useful in places and for purposes its inventors did not anticipate.
For Zhuangzi, even the legendary Liezi, who can travel by riding the wind, remains dependent on it. That does not make his ability worthless. It changes what freedom means. Open weights remove one gatekeeper, not every condition of existence; building the data centers gives another kind of power to those who finance and operate them.
Peng also has a destination: the southern sea. Our discussion of AI is richer in measurements of wings and wind than in agreement about where such intelligence ought to go. The capacity to act is not the wisdom to choose an action.
Alibaba’s poster has left the fish’s name blank, though any reader of Zhuangzi can fill it in. The more difficult sentence is the one we have yet to write: who will control the conditions of flight, and to what end?
Noman Shah is fellow of the British Computer Society, founder of a UK sovereign-AI company and author of ‘How DeepSeek Unleashed Pandora’s Box.’
