Something strange happened in July 2026. A small company in Beijing called Moonshot AI built one of the best AI models in the world. Naturally, it cost a huge amount of money to create. However, the company then put the whole thing online and let anyone download it. For free. The model is called Kimi K3. Furthermore, it is very good at coding and writing. In fact, it beats several AI models that big American companies charge a lot of money for. Meanwhile, this was not a one time event. Chinese companies have been doing the same thing again and again for eighteen months. For example, DeepSeek did it. Similarly, Alibaba does it constantly, and so do Z.ai, MiniMax, Tencent and Xiaomi. So here is the question everyone keeps asking. Why would anyone spend millions of dollars building something, and then give it away? The answer is not kindness. Instead, it is a very smart business plan. Above all, it is working.

What “Free” Really Means With China Open Source AI Models

First, you need to understand one small detail. Otherwise the rest of the story will not make sense, because most news reports get this part wrong.

What is a “weight” in an AI model

An AI model is really just a huge list of numbers. Those numbers are called weights. Think of it like a recipe with trillions of tiny measurements inside. During training, all those numbers get adjusted slowly until the AI can write code and answer questions properly.

Closed models versus open models

When you use ChatGPT or Claude, for example, you never get those numbers. Instead, the company keeps them locked on its own computers. You type a question, their machine does the work, and an answer comes back. In other words, you are renting. Nothing ever belongs to you. Now compare that with an open weight model from China. Here, the company hands you the actual numbers. As a result, you can download the file and put it on your own computer or server. Because it runs on your hardware, nobody sees your data and no monthly bill arrives. In short, you own it.

Open weight is not the same as open source

These two terms sound identical, but they are not. Open weight simply means you get the finished model. Open source, on the other hand, means you also get the building instructions plus details about the training data. James Landay, a computer science professor at Stanford, says people mix up these two words all the time. Indeed, most Chinese models are open weight only. To put it another way, you get the cake, but never the recipe. Remember this point, because it matters later when we reach the risks.

How Good Are These Free AI Models

Very good. That is exactly what shocked people. Kimi K3 has 2.8 trillion parameters, which makes it the biggest free model ever released. Moreover, it can read one million words at once. Because of that, you can hand it an entire codebase or a full book without splitting anything up. A testing company called Artificial Analysis checked the model independently. Their result placed it fourth out of 189 models worldwide. In other words, only two American models scored higher. It also finished first on a leaderboard for front end coding, where real developers compare answers without knowing which AI wrote them.

And it costs much less

Here is the part that worries American companies.

• Kimi K3 costs about $3 for a million words of input
• OpenAI’s top model costs about $5
• Anthropic’s top model costs about $10

Look at the cost of finishing a full task and the gap grows wider. K3 landed near $0.94, while a top American model landed near $1.80. That is roughly half price. Demand became so heavy that Moonshot stopped taking new customers three days after launch. Simply put, the company did not own enough computers. Surprisingly, that last detail is the key to this entire story.

Why China Is Giving Away Its Best AI Models: 8 Simple Reasons

Reason 1: China cannot buy enough computer chips

Since 2022, America has blocked sales of its best AI chips to China. Chinese companies still have brilliant engineers and plenty of money. What they lack, however, is computing power. That single shortage changes everything. If you cannot buy chips, then you cannot serve millions of users yourself. Eventually you run out of capacity. Indeed, that is exactly what happened to Moonshot within three days. Kyle Chan, a researcher at the Brookings Institution who studies Chinese tech policy, explains that this shortage makes the free strategy far more attractive. Chinese labs, in fact, complain about it constantly. So they flipped the problem around. If you cannot serve the world yourself, then let the world serve itself.

Reason 2: Free models borrow other people’s computers

This is the clever part. The moment Moonshot published the file, every cloud company on earth could start running it. Databricks, Together AI, OpenRouter and Modal all jumped in. Basically, any business with a room full of graphics cards became a delivery channel. None of them pay Moonshot anything. Even so, all of them spread the model further. Chan calls this an amplifying effect. Other companies already spent billions on hardware, and now that hardware quietly delivers a Chinese model instead. Certainly, no advertising budget can buy that.

Reason 3: Free is the best advertising ever invented

Two years ago, almost nobody outside China had heard of Moonshot AI. Today, its model sits on the front page of every tech site. Whenever you give away something your rival sells, developers love you for it. And developers matter more than most people realise. First they pick the tools. Then those tools become part of the product. Finally, the product becomes company standard, and switching is nearly impossible. Consider Alibaba as proof. Its Qwen models passed one billion downloads on Hugging Face by March 2026, faster than any AI family in history. Meanwhile, users have built more than 113,000 modified versions. That is not just a product anymore. Rather, it is an entire world built around one company.

Reason 4: Make your rival’s product feel worthless

There is an old trick in technology. Whenever your rival sells something, make it free.American AI companies have one main product, and that product is access to their models. Now imagine a free model that is almost as good. Maybe 85 or 90 percent as good. Suddenly the expensive option must justify its price every single month. Actually, this is already happening. Around 60 percent of companies have capped their AI spending. As a result, many now mix both. Easy jobs go to the cheap free model, while only the hard jobs go to the expensive one. According to a RAND study, Chinese models can cost as little as one sixth of American ones. Clearly, that is very hard to compete against.

Reason 5: Beijing turned free AI into foreign policy

This part is much bigger than any single company. On July 17, 2026, at a large AI conference in Shanghai, President Xi Jinping told the audience that AI should not be controlled by one nation. Specifically, he called for “open source, collaboration and sharing.” Additionally, he announced an AI alliance covering 29 countries. This message also appears in official policy. Chinese government plans mention open source as a national goal. Notably, one document argues that AI must not become a game only for rich countries and rich people. State media repeats the same idea. Free Chinese models, they argue, let poorer countries build their own AI rather than depend on costly American systems. And the pitch is landing. Malaysia, for instance, has explored using DeepSeek as a national model on Chinese chips. Across many developing countries, Chinese models are now the default choice, mainly because they are the only powerful option a normal government budget can actually own.

Reason 6: Whoever is free becomes the standard

History keeps teaching this lesson. Linux was free and it beat expensive business software. Likewise, Android was free and it took over most of the world’s phones. Neither one won by being perfect. They won because everybody could use them. Now think about the next million AI apps. If most of them are built on Qwen, DeepSeek or Kimi, then Chinese design choices quietly become the world’s default settings. A US government report described exactly that pattern. Chinese labs have already invented technical methods that the rest of the industry now copies. Ultimately, owning the foundation lasts far longer than winning one product launch.

Reason 7: Free models attract the best scientists

Top AI researchers want their work to be used, not locked in a vault. Suppose you had two job offers. In the first, your work sits behind a paywall. In the second, a hundred thousand people download and build on it. Most people, understandably, pick the second one. Chinese labs know this, and they use it to hire talent.

Reason 8: They still make money anyway

Free weights do not mean zero income. Moonshot still sells its app and its API. Similarly, Alibaba sells cloud services and chips, while Z.ai sells coding tools. It is the same trick Red Hat used with Linux years ago. Give away the main thing, then charge for the convenience, the support and the setup. Here is the funny part, though. Most people who download a 2.8 trillion parameter model soon discover they cannot run it, since it is far too big for a normal computer. Consequently, many of them come back and pay for the hosted version anyway. To sum up, the free model is the front door, not the whole house.

The Numbers Behind China Open Source AI Models

This is not just talk, because the data is very clear.

• Downloads: During the year ending August 2025, Chinese free models took 17.1 percent of global downloads on Hugging Face, while American models took 15.86 percent. That was the first time China led. By spring 2026, moreover, the Chinese share had climbed to roughly 41 percent.

• Popular models: In recent weeks, the six most used models on OpenRouter were all free Chinese models from Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai. Notably, the first American model sat in seventh place.

• Real traffic: On Vercel, which powers many websites, free models handled almost one third of all AI requests in June 2026.

• Speed of growth: According to RAND, traffic to Chinese AI models jumped 460 percent in two months after DeepSeek launched R1. During that time, their global share rose from 3 percent to 13 percent.

• Size of the ecosystem: Hugging Face now hosts nearly three million public models, and a new one appears roughly every seven seconds. Unsurprisingly, Chinese models sit at the centre of that activity.

American companies quietly use these models too. For example, Airbnb’s CEO confirmed the company uses Qwen for AI customer service, while Pinterest has been testing open models. Many small startups switched simply because their AI bills grew too large.

Who Is Behind These Free Models

DeepSeek

The company that started everything with R1 in January 2025. Since then, it has become famous for cheap models that punch above their weight.

Alibaba Qwen

Currently the biggest free AI ecosystem in the world. In addition, Alibaba builds its own chips so these models can run on Chinese hardware.

Moonshot AI

Maker of Kimi. At present, K3 is the largest free model ever released.

Z.ai

Maker of the GLM models. Interestingly, when Hugging Face faced a security incident, American models refused to help because of their safety rules. Hugging Face therefore used Z.ai’s model instead.

MiniMax, Tencent, ByteDance and Xiaomi

All of them release strong models. ByteDance in particular is enormous inside China.

How American Companies Are Reacting

Not calmly. Dean W. Ball, who works on strategy at OpenAI, wrote a post warning that a world full of free AI models could become a nightmare. Specifically, he argued that free models discourage companies from spending big on research. Later, however, he backed away from part of that claim after heavy pushback from other well known tech figures. Anthropic raises a different worry, and its concern is safety. Once a model file is released, after all, you cannot take it back. Nor can you fix its safety rules later or stop someone from misusing it.

Washington is moving as well. Reportedly, the Trump administration wants to restrict Chinese AI models for security reasons. Congress has already questioned American companies about using them. Furthermore, one consulting firm compared the risk to the old Huawei debate. Plenty of people disagree, though. David Sacks, a White House AI adviser, warned publicly that a ban would simply hand the market to two big American firms and hurt innovation. Many tech companies signed a letter against broad restrictions. Researchers add that losing access would damage science, since you can study a free model in ways a locked one never allows.

There is also a practical problem. You cannot delete a file that millions of people already downloaded. Therefore any ban would only affect the future, never the past. America does have its own free models. OpenAI released some in 2025, Nvidia released one in March 2026, and a startup called Thinking Machines Labs released one in July 2026. Even so, China remains far ahead in this area.

The Risks You Should Know About

Admittedly, this article is not saying Chinese AI is perfect. There are real concerns worth checking.

Censorship is built in

These models are trained under Chinese content rules. As a result, they will avoid or dodge certain political and historical topics. If you are building a news app or a research tool, therefore, test this carefully before launch.

You do not know what is inside

Getting the model file does not mean getting the full story, because you usually never see the training data. Professor Landay warns that you may not know what sits inside the model or how it behaves with your information.

Code safety

The main worry security experts raise is that a model could quietly write code with weaknesses in it. Running it on your own server protects your data. Nevertheless, it does not remove the need to review the code it produces.

The copying argument

American labs accuse Chinese companies of training on their models’ answers, a practice known as distillation. Anthropic, for instance, said it found thousands of fake accounts created for this purpose. Then in July 2026 a White House official accused Moonshot directly, and sanctions were mentioned. Moonshot has not admitted anything, however. Some experts also point out that the timeline does not fully add up. Consequently, treat this as an ongoing argument rather than a proven fact.

The Twist: China Might Close the Door

Here is the part nobody expected.

On July 7, 2026, Reuters reported that China’s Ministry of Commerce had been meeting with Alibaba, ByteDance and Z.ai. Reportedly, the topic was whether to limit foreign access to China’s best AI models, including the free ones. Two weeks later, on July 21, the Financial Times reported that China is also weighing tighter export rules on AI and chip technology. Meanwhile, a group of Chinese legal experts suggested a system with three levels. Basic tools would only need registration. Medium tools would face a security check. Finally, the most powerful models would either stay inside China or never be released publicly. Nothing has been decided yet, since there is no draft law and no date. Any rules might also apply only to future models. Still, the direction is clear. Both countries are starting to treat AI models the way they treat weapons technology. In other words, this moment, where some of the world’s best AI is free to download, may not last.

What This Means for You

If you build software: right now you hold real power. Because near top level AI costs a fraction of the usual price, you can also run it yourself. Test a free model on your own real work before assuming the expensive one wins. In many cases, teams find the cheap option handles most jobs fine.

If you run a business: do not pick one side. Instead, send high volume simple work to cheap free models, and save the expensive model for hard problems. Additionally, keep a list of what you depend on, because the rules could change quickly.

If you live in a country with a small tech budget: this shift matters enormously. Owning a powerful AI, running on your own machines and in your own language, went from impossible to affordable in roughly eighteen months.

If you are simply curious: notice that the AI race is no longer only about who builds the smartest model. Rather, it is about whose model everything else gets built on. That is a very different race, and China currently leads it.

Frequently Asked Questions

1) Why does China give away its AI models for free?

Mainly because chip restrictions mean Chinese companies cannot serve millions of users themselves. By giving the model away, they let other people’s computers do the work instead. Additionally, the strategy wins over developers, pressures American prices, and supports China’s goal of an AI world not controlled by one country. Meanwhile, the companies still earn money from apps, cloud services and business deals.

2) Are Chinese AI models as good as ChatGPT or Claude?

Close, although not quite equal at the very top. In July 2026, for example, independent testing placed Kimi K3 fourth in the world. It beat several expensive American models, yet it lost to the best two. For everyday tasks you probably would not notice a difference. For the hardest work, however, the top American models still lead.

3) Is it safe to use Chinese AI models in a business?

That depends on the job. Running the model on your own server protects your data. Even so, it does not solve concerns about built in censorship, unknown training data or code quality. If you work in a regulated industry, therefore, run your own security review first.

4) What is the difference between open weight and open source?

Open weight means you receive the finished model file. Open source, by contrast, means you also receive the instructions and data details used to build it. Most Chinese releases are open weight only.

5) Will China stop releasing free models?

Possibly. During July 2026, both Reuters and the Financial Times reported that China is considering limits on foreign access. Nothing is final yet, and any rules may apply only to future models. Meanwhile, files that are already public can never be taken back.

6) Can I run these models on my laptop?

The biggest ones, no. Kimi K3 needs serious server hardware. Fortunately, smaller versions of Chinese models do run on a normal computer, and many people use them that way every day.