Qwen, Alibaba's Chinese open-source AI model, has become the world's most downloaded artificial intelligence with 700 million downloads - and is quietly powering a growing number of Western startups. Why the silence and what does it mean for you?
While you were wondering whether ChatGPT or Gemini, Silicon Valley has already secretly moved on to something entirely different. It's called Qwen, hailing from Hangzhou, has just become the most-transmitted artificial intelligence on the planet. And no, most of the West has never heard of it.
Imagine America spending a few hundred billion dollars, building the largest data centers in human history, elevating Nvidia to the most valuable company in the world, and then banning the export of its best chips to China just in case—to maintain its advantage. And then one morning discovering that half of its most promising startups are quietly running their products on the Chinese model, which is free.
This is not a conspiracy theory from YouTube. This is a snapshot of the state of affairs at the beginning of 2026.
How a silent revolution happens
The story begins in January 2025, when China's DeepSeek released the R1 model. It was open source, free to download and modify, and reached a level of sophistication previously thought to be reserved for expensive, closed systems from California. The reaction in Silicon Valley was somewhere between shock and panic. The stock market wiped out several hundred billion dollars in value from AI-related companies in a single day.
But R1 wasn't a one-off event. It was the starting shot.
Then came Qwen, Alibaba's family of models. And that's where the story really comes to life. According to data from the Hugging Face platform, Qwen had reached 700 million downloads, making it the most popular open-source AI system in the world. It overtook Meta's Llama family in cumulative downloads as early as October 2025.
The number that really tells the story, however, is not 700 million. In December 2025, Qwen amassed more downloads in a single month than the next eight most popular models combined—including Meta, DeepSeek, OpenAI, Mistral, Nvidia, Zhipu, Moonshot, and MiniMax. This is not a close race. This is a sweep.
And the scale is absurd: Alibaba has open-sourced nearly 400 models from the Qwen family, which have spawned over 180,000 derivatives and support 119 languages and dialects.
When 80 % is not what you think
Here we come to the claim that's been circulating the web in all caps: "80 percent of American startups have switched to Chinese AI." It sounds dramatic. And it is - but there's one important star you need to know before you start throwing ashes or getting excited.
The number comes from Andreessen Horowitz (a16z), one of the Valley’s most powerful venture funds. Partner Martin Casado said that when entrepreneurs come to a pitch today, the probability that their startup is running on a Chinese open-source model is very high — he estimates it at about 80 percent. His colleague Anjney Midha was even more blunt: “It’s really a Chinese game right now.”
An important detail: this 80 percent refers to companies that build on open source tools – not all American companies in a row. The difference is not small, and it is fair to say it out loud, otherwise we will remain with a cliché headline instead of the truth. And yet the trend is absolutely real: on the OpenRouter platform, which tracks the actual use of models among developers, Chinese open models grew from almost zero market share at the end of 2024 to almost 30 percent.
The change happened quickly. Faster than most noticed.
Mathematics is brutal
Why? Because developers aren't sentimental. They're accountants.
According to a16z analysis, startups using DeepSeek-class models pay about $0.10-0.20 per million tokens. The same workload on leading closed models costs $20-60 per million tokens. That’s not a small difference. It’s a different planet.
The specific prices tell the same story: Kimi K2 charges just 15 cents per million input tokens and $2.50 per million output tokens, while Alibaba's Qwen3-Max starts at $0.459 per million input tokens. Overall, the price differences between Chinese and American closed-loop models reach a factor of 10 to 40.
For an early-stage startup that looks at how much money it has left each month, this isn't a philosophical question. It's the difference between whether the company survives next year or not.
And another breathtaking detail: According to the company, DeepSeek-V3 was trained with about $5.5 million and 2,000 Nvidia H800 graphics cards – a cheaper version aimed at the Chinese market – yet it reached the level of leading closed-loop models costing hundreds of millions.
The secret your program hides
The best part of this story is its discreetness. Many Western products already run on Chinese foundations – they just don't advertise it very loudly.
Take the popular programming tool Cursor. In March 2026, a user on the X network noticed suspicious traces in the code of a new model, Composer 2. Cursor executives later admitted that Composer 2 was initially built on top of Moonshot AI’s open-source Kimi K2.5 model. Co-founder Aman Sanger candidly said, “We made a mistake by not mentioning the Kimi framework in the announcement from the beginning.” According to the explanation, about a quarter of the total computing power used to build Composer 2 came from the Kimi framework, and the remaining three quarters came from Cursor’s own continuous learning.
And he's not the only one. Cognition, the company behind the Devin programming agent, has unveiled the SWE-1.5 model, which is believed to be based on Zhipu's (Z.ai) GLM model.
The industry is no different. Thinking Machines Lab, a $12 billion startup led by Mira Murati, the former CTO of OpenAI, has released a tool for adapting existing open-source models—including eight from the Qwen family.
Even Airbnb didn't hold back. CEO Brian Chesky described Qwen as "very good, fast, and cheap." In short: a model you can't really complain about.
So why the silence? Because in the context of the US-China technology rivalry, the statement “our AI uses the Chinese model at its core” is not just a technical detail for an American company, but also an exposure to public relations risk. How will investors see it? Will corporate customers be concerned about data security? What headlines will the press write? It’s easier to just say nothing.
When sanctions turn against you
This is where the story takes on a real flavor of irony that a screenwriter could write.
Washington has restricted exports of Nvidia’s most advanced chips to China in an effort to slow its development. The result? Without access to cutting-edge chips, open-source publishing of designs has become a shortcut for Chinese labs—accelerating external feedback and contributions that make up for limited computing power. In other words, the restriction has forced them to become more efficient and more open, which is exactly what the West wanted the least.
At the same time, the sanctions have ignited a domestic chip industry. Tightening U.S. controls on exports of advanced Nvidia chips have fueled the boom by boosting demand for domestic alternatives. One analyst summed up the situation by noting that Nvidia’s once-dominant position in China’s chip market will all but evaporate by 2025.
And the money poured in. Hong Kong became the world’s number one IPO market in the first quarter of 2026, with 40 listings raising about $14 billion, more than the Nasdaq, New York Stock Exchange and Bombay Stock Exchange combined. The year’s top performers were Zhipu and MiniMax, which jumped more than 400 percent after their IPOs. AI chip designer Biren Technology closed 76 percent higher on its Hong Kong debut, raising about $707 million.
A lesson for anyone who thinks progress stops with a single ban: water always finds a way.
Even a fairy tale has fine print
Make no mistake – this is not a free lunch. The open source model is free, but the infrastructure to power it at scale is not. Alibaba CEO Joe Tsai put the strategy bluntly: opening up Qwen models “democratizes AI” and “expands applications” – which directly feeds their cloud computing business. More developers using Qwen means more companies needing Alibaba’s servers. The model is the bait. The servers are the hook.
There’s also the question of data: while Western startups build on these open models, their actual use helps refine the next wave of Chinese systems. Whoever controls the foundation has a quiet but real influence on how the rest of the world builds.
And the human element is not insignificant. In March 2026, Qwen Family’s lead engineer, 32-year-old Lin Junyang, posted a short message about his departure, which alarmed the developer community and sent Alibaba’s shares down 5.3 percent. Even invincible giants are actually made up of people who simply leave one morning.
The end of carefree monopoly
So what's left after we subtract the panic, the clickbait headlines, and the geopolitical drama?
It remains a simple, almost childish realization: AI has become a commodity. Just last year, the West’s advantage seemed like an insurmountable moat—hundreds of billions of dollars, the best chips, the best teams. Today, Chinese labs offer about 90 percent of the capabilities of top-of-the-line models for a fraction of the price, and they do so openly, without questions, and without asking for permission. It’s a story that should be known not just by engineers, but by anyone who runs a business, invests, or just wants to understand where the world is going.
Let me be honest: I see something very healthy in this. Monopoly is never good for the user, and competition lowers prices and forces everyone to do better. The fact that today a small team in Ljubljana, Nairobi or São Paulo can build a serious product for a few cents per million tokens instead of a fortune is downright liberating. Just two years ago, this would have been science fiction.
But caution is in order. Free tools that become the foundation of your business come with hidden costs – dependency, data issues, and geopolitical uncertainty that you don’t control. So smart companies do what any sensible person should: they don’t bet everything on one card. They use premium Western models for the most sensitive tasks, and Chinese open-source models for everything else where speed and price matter.
So the best advice I can give is old-fashioned: know what your product is actually running on. Because in 2026, it's entirely possible that your favorite software utility is thinking in Chinese in the background—and just doesn't want to tell you.
