China is surpassing the Western AI industry with a new neural network that is "more powerful than DeepSeek."

Recently, China shook the world by launching the neural network DeepSeek. The power of this new product, developed for a mere $12 million compared to the hundreds of millions spent by American competitors, left a grim impression on Silicon Valley. Stocks of Western AI companies plummeted, with chipmaker Nvidia losing $600 billion in an instant. But misfortunes (for the West) do not come alone. On the first day of the Chinese New Year, major AI player Alibaba presented the Qwen 2.5 Max neural network, which is even more powerful. At this rate, China’s leadership is becoming indisputable.

NOT FROM SCRATCH

Alibaba is to China what OpenAI is to the USA; the company has long and successfully developed networks. Last spring, a new company, DeepSeek, burst onto the technological horizon of the Middle Kingdom (founded in 2023). The first version of their neural network went unnoticed in the West (and now they are regretting it). However, it made quite a stir in China, prompting Alibaba to hold meetings and crisis management sessions. The youth are outpacing us!

And now there’s a new twist. The network from Alibaba is breaking into the domestic market, and together with DeepSeek, they are effectively carving up the international market.

QWEN 2.5 MAX: WHAT IS IT?

The product from Alibaba outperformed DeepSeek in all the benchmarks typically used to evaluate neural networks. The network excelled in answer accuracy, data processing speed, and adaptability. This is not surprising: it was trained thoroughly.

In reality, the creators of neural networks have long been competing primarily in training. According to American OpenAI, this process is very expensive and time-consuming. For instance, OpenAI plans to spend a billion on training the next generation GPT, GPT-5. Companies are acquiring entire power stations to fuel their training centers. Additionally, content theft from various sources (the internet, social media, even personal data) has long been a subject of disputes and investigations. However, content is critically needed, as AI requires examples to understand human life.

DeepSeek became a significant neural network because it managed to train itself for just $5.5 million (though some expenses are reportedly not included in this total; still, it’s inexpensive), and the results are quite impressive. The neural network is certainly no worse than GPT-4 but cost hundreds of times less.

Qwen was not trained on a shoestring budget. It was loaded with 20 trillion units of information (tokens) and additionally employed methodologies like fine-tuning (SFT) and human feedback (RLHF), meaning the most modern techniques were applied. The process may not be as fast as with DeepSeek, but it will be better. It’s no wonder that Qwen surpassed existing American counterparts.

Alibaba does not claim that it trained the network on the cheap. However, it suggests that it was not as expensive as for the Americans. The company provided detailed insights into its training methods. What struck experts the most was the optimization at every stage. As a result, the minimum computational power is required for both training and operation.

HOW TO USE QWEN 2.5 MAX

Reports indicate that Qwen can process 1 million units of information at once. This is akin to being able to hold, word-for-word, 10 novels the size of “The Lord of the Rings” in your head at any moment. Experts say this is a turning point in AI development. Most neural networks can barely remember 120,000 units of information.

Developers are already rushing to test it: the network is available for experimentation under a free license.

Although Qwen, like DeepSeek, has a chat feature allowing users to converse with the neural network (after registration), primarily, Qwen is a toolkit from which advanced users can craft personalized solutions, such as creating images, analyzing computer codes, solving engineering and scientific problems, and so on.

In fact, this is how the GPT we know operates. You have likely sought help from GPT, but where did you find it? You didn’t download it yourself, but rather accessed it through a platform like Telegram or a specific website. This means someone purchased access to the main system, adapted it, and now offers it to average users. Their interest typically lies in advertising, which is inserted into the Telegram channel or website, or in the copyrights for the AI-generated images and texts.

Feedback from Russian users is still limited, while the innovation has been well-received in the West.

A BRAVE NEW WORLD

The launch of DeepSeek has drawn the attention of the U.S. President: Donald Trump stated that global competition is intensifying, and American companies need to step up their game to keep ahead.

Trump's team takes AI technologies very seriously. They understand that whoever possesses powerful AI controls the modern world. AI in the hands of your enemy reveals all your secrets, laid bare. Not to mention that powerful AI can lead to rapid development of new medicines, technology, and weaponry.

What can America do? Experts agree that American companies find themselves in a less than ideal situation. First, they were pioneers and spent a fortune on what is now “obvious.” It’s like with mobile communications: the first phones and tariffs are expensive. Secondly, large companies have become bloated and lost momentum. Huge expenditures on salaries and non-essential projects, prioritizing management over engineering – it’s the same old story.

Can the West rally a cohort of young, driven startups and engage in a race with new forces? Most likely, yes, but they will have to go through a crisis.

In any case, 2025 is indeed becoming the year of AI, but not because OpenAI is releasing some astonishing foundational neural network. It’s far more interesting than that.