Chinese AI models are gaining popularity, including in the West. The models often perform at the level of more expensive LLMs from American companies. However, they don’t always achieve that efficiency in a kosher way.
The most widely used AI model at the moment is the Chinese Deepseek, according to figures from OpenRouter. That company offers a unified interface for LLMs, through which customers can choose which model they use. OpenRouter by no means provides a complete picture of AI usage worldwide, but it can provide concrete figures to show how users, when given the choice, quickly turn to Chinese models.
Deepseek is, after all, not the only popular LLM provider. In second and third place are the LLMs from Google and OpenAI, but Qwen is nipping at Anthropic’s heels in fourth place. Looking at specific models based on usage, Deepseek V4 Flash is the most popular, followed by LLMs from Xiaomi, Tencent, and Z-AI. In the Benelux as well, companies with AI experience are starting to wonder which LLM is most suitable for a given task, and they are increasingly ending up with Chinese alternatives.
A powerful start
This should come as no surprise. When the Chinese company Deepseek launched Deepseek R1 in early 2025, the AI world was shaken to its foundations. Against all expectations, the Chinese had built a model that could compete with OpenAI’s LLMs, and they did so without access to advanced (and expensive) Nvidia accelerators for training. Furthermore, Deepseek R1 was more efficient than its American competitors.
That trend has continued since then. Anthropic, OpenAI, and Google remain the absolute AI champions. The American companies regularly launch new models that are almost alternately the most powerful in the world. No other company in the world can currently release an LLM that will suddenly surpass Fable 5 or GPT-5.6.
No race for gold
That is not what the Chinese are after either. Not long after, say, Anthropic or OpenAI stuns with a new Frontier model, companies like Deepseek or Z-AI come out with an LLM that is slightly less good, but still very powerful. That LLM is then more than good enough for numerous tasks. Additionally, Chinese models are generally more efficient than the large Frontier LLMs. They use fewer tokens and therefore require less computing power.
Furthermore, Chinese companies are bringing their models to market under open-weight licenses, making them free to use. This lowers the barrier for organizations worldwide to integrate Chinese LLMs. When Deepseek R1 was launched in 2025, for example, Microsoft was quick to make the model available for Copilot+ PCs immediately.
Chinese models thus have many characteristics that make them attractive. They are more efficient than Frontier models, and more open so that users create less dependency on a single player. Companies that have their AI strategy in order and can easily switch between LLMs are understandably attracted to the cheaper alternatives. Why pay more for a Google model for simple prompts or automated workflows if one from Z-AI can do the same?
Trillion-dollar valuation at risk
This trend is also noticeable in the US, and the big AI companies don’t like it. They see poachers on the horizon. OpenAI and Anthropic both have an IPO in mind where they want to be valued at at least a trillion dollars. That valuation is hard to justify when their primary products are consistently challenged by cheaper alternatives within a few months.
Furthermore, the use of AI is increasingly evolving into a system of frameworks, where organizations build harnesses and context systems. These then provide access to data and integration into workflows, where the brain driving a task can be easily swapped. GPT-5.6 Luna may be preferred to drive an agent at launch, but switching to a Deepseek variant that turns out to be cheaper two months later is no longer a major hurdle.
Angry at openness
A bit of lobbying comes as no surprise in this context. In the US, voices are being raised to ban open AI (not OpenAI). Unspecified security reasons are being used as a stick to beat it with. Open models are far from exclusively Chinese, but a large share of the popular open-weight LLMs do originate from China. The goal of such a ban is therefore clear.
Meanwhile, many (American) tech giants have united in the Open Secure AI Alliance. With this alliance, they want to safeguard open AI. Notable absentees from the club: Anthropic and OpenAI.
Understandable suspicion
Still, some suspicion toward Chinese AI models is understandable. After all, how is it possible that the Chinese can deliver such good AI models so quickly without access to the most advanced hardware for training?
The Chinese government claims that Chinese companies are simply doing good work. AI researchers are said not to be stealing knowledge from the Americans. On the contrary: companies in the US are allegedly trying to reproduce Chinese models.
Sneaky distillation
That claim is at odds with new and extensive research from Reuters. Reuters examined more than 80 academic papers and patent applications and, based on documents, determined that Chinese researchers linked to the government are indeed guilty of what is called model distillation.
Model distillation refers to a technique where a closed LLM is used to train new AI models. An LLM such as Anthropic’s Fable 5 is then used as a kind of teacher that trains new, smaller models. These distilled student models are never as powerful as the original, but when the distillation focuses on specific capabilities, they can perform them almost as well and more efficiently.
The US has long accused China of distillation, and the Reuters investigation now seems to provide fairly conclusive evidence. This is not to say that all Chinese AI companies only distill models, but the practice seems to play a role in China’s rise as an AI powerhouse.
That explains to some extent where the efficient Chinese AI models come from. Are the voices calling for a ban on open LLMs right then?
Distilled, not copied
That is still a different question. After all, distilled models are not copies. They are LLMs with a right to exist, as their efficiency proves. If OpenAI were to distill its own LLMs into new models, there would be no problem at all.
In this case, however, Chinese companies and researchers are using American models to secretly distill their own LLMs without permission. According to the owners of those closed models, that is intellectual property theft.
The pot calling the kettle black
Anyone who choked on their coffee at that last sentence is forgiven. After all, haven’t OpenAI, Anthropic, and Google trained their models by virtue of a massive theft of intellectually protected property that served as training data? GPT-5.6 can only write well because it has read and used almost all books by all authors as training material without permission. Now that GPT-5.6 itself is being used uninvited to train AI, there is suddenly talk of unauthorized IP.
If the rise of Chinese AI models tells us anything, it is perhaps primarily that the development of expensive Frontier models is no guarantee of a large return on investment. The initial investment is enormous, but through the technique of distillation, it is possible to build more efficient specialized models relatively quickly.
Useful models
In this regard, we note that distillation can never result in a more powerful LLM. Fable, Mythos, and Gemini remain the absolute champions. However, that is not so relevant within a business context. An AI agent does not need to work with a super-powerful LLM. On the contrary: the more compact and efficient, and the lower the token cost, the better.
Dismissing Chinese models as a cheap copy is therefore certainly not correct. That sneaky distillation is problematic from an intellectual property standpoint hopefully needs no further argument. On the other hand, it shows a certain hypocrisy to make a problem of IP theft now, and not during the initial gathering of training data.
Helpful
For companies, the Chinese are currently delivering interesting LLMs. Furthermore, the open-weight models contribute more to the public development of AI than the closed models from OpenAI, Google, and Anthropic.
LLMs from the US are increasingly subject to American regulations and could suddenly become unavailable if the US president gets out of bed on the wrong side. From a European perspective, Chinese LLMs therefore seem more helpful than problematic at the moment.