Have you heard of Qwen?
When I wrote my first article on the economic effects of industrial robotics in a Spanish newspaper (I think I was seventeen at the time), and later, while discussing my doctoral thesis in San Francisco with Professors David Mowery and Laura D’Andrea Tyson at the University of California, Berkeley, I soon realized that the great technological paradigms (for that is what they must be called) manifest themselves through major technological trajectories (which must also be called as such) in order to lead the diffusion of innovations within the economic structure.
We spent endless hours constructing econometric mathematical functions of diffusion — the epidemic model — often simple logistic functions that we tried to fit into data series like Cinderella’s slipper after midnight.
On weekends I devoted my time to visiting the missions founded by Spaniards throughout California. One of them was the Mission Santa Clara de Asís, founded by Spaniards in 1777, which had by then already become the nerve center of Silicon Valley. I thought of it again a few years ago when I visited Hangzhou, the city that hosts the company Alibaba — a city which, at the end of the thirteenth century, Marco Polo described as “the most splendid and elegant city in the world.” And what does one have to do with the other?
Since the steam engine, probably nothing has fundamentally changed in the processes of diffusion and in the development of trajectories following the emergence of paradigms. The bet is risky for investors. If they choose the wrong trajectory, they may end up in the cemetery. The first video cassette recorder (VCR) commercially available was the U-matic system, which appeared in the autumn of 1971. From that moment onward, a fierce competition unfolded among incompatible formats and licensing systems for analog video recorders and players: Sony’s Betamax, JVC’s VHS, and Philips’ Video 2000. VHS won (as you know). The others may be visited in the cemetery.
The race for artificial intelligence is not confined solely to the United States. It is true that in San José (primarily) the most advanced models have been developed. Companies such as OpenAI (San Francisco), Google (Santa Clara), Meta (San Mateo), Microsoft (Seattle), and Anthropic (San Francisco) have produced the well-known models ChatGPT, Gemini, Claude, and LLaMA.
Americans believe in private innovation; they rely on startups with brilliant storytelling, play the game of venture capital, and are supported by the best universities. There should be no doubt: in advanced research the American models — GPT, Claude, or Gemini — are ahead. This includes the chips (GPUs) produced by Nvidia (Santa Clara, USA), the true engine of modern artificial intelligence, along with the work of AMD (Santa Clara, USA), Intel (Santa Clara, USA), and TSMC (Hsinchu, Taiwan).
In China, where now I live, we work in other spheres, with companies such as Alibaba, Baidu, Tencent, and ByteDance (TikTok). In due course we discovered that their models are extremely useful: Qwen from Alibaba, Ernie from Baidu, Doubao from ByteDance, and Hunyuan from Tencent.
In the West there is a widespread belief that China’s advantage rests solely on state planning, government support, and coordination among companies. Yet although the state of the People’s Republic has turned AI into one of its principal national strategies, among state-owned enterprises there exists a curious coordination — and at times a fierce competition. In basic AI research China is indeed behind California, yes, but…
In large-scale application China is unbeatable; it is far ahead of the United States. In addition to an admirable research effort, it possesses a larger universe of users (population), access to a greater volume of data, and far fewer regulatory restrictions. China’s capacity for experimentation through trial and error is extraordinary. As far as I know, China already has more than 200 registered generative AI models; it is the country with the largest number of AI users in the world and has the greatest number of software engineers on the planet.
So who is winning? Beta or VHS?
Qwen is a family of generative artificial intelligence models similar to ChatGPT or GPT-4, but developed in China. In a conversation with one of the most senior executives at Alibaba, he explained to me the forthcoming launch of JVS Claw in order to help users install OpenClaw (you will be hearing about this in the coming months). Users, I thought — the key to China’s advantage.
In April 2023, Alibaba launched the first beta version of Qwen, calling it Tongyi Qianwen. While attending a conference in Beijing in September 2023, I witnessed the announcement of its public release. Three months later, during the Christmas period of 2023–2024, I was informed that they were launching the open-source models 1.8B and 72B. From then until today, Qwen’s growth has been staggering.
In 2024 (which already feels like a long time ago), Alibaba launched Qwen2, introducing Mixture-of-Experts models. The real surprise, however, was QwQ, a model specifically designed for complex reasoning tasks. At the beginning of 2025 came Qwen2.5-Max, Qwen2.5-VL, and Qwen2.5-Omni, enabling Alibaba to generate images, video, and audio through its own language models. That same year, Qwen3 was already being trained on 36 trillion tokens across 119 languages, paving the way for Qwen3-Max, Qwen3-Omni, and Qwen3.5.
The reader of this article may conclude that the evolutionary capacity of this model is remarkable. Yet this is not the most striking aspect. What is truly significant is that Qwen’s open weights allow them to be modified, retrained, and integrated into other applications. Its diffusion resembles a genuine supersonic rocket.
The architecture itself is highly efficient. The Mixture-of-Experts approach relies on specialized neural networks that are activated only when required, resulting not only in lower computational costs but also in greater scalability.
The diffusion capacity of Qwen — open and modifiable — is extraordinary. Who will ultimately prevail? The United States or China, GPT or Qwen, chips or data? What matters is not who will win, but how much we will all gain. Companies are already casting a sideways glance at Qwen: if it is cheaper, more flexible, and better, what exactly are we waiting for?
Santa Clara or Hangzhou?
Si (
No(
