DeepSeek and Qwen: How China Turned Open Models into a Leap in the Global AI Race

Chinese open source in AI has moved past catch-up mode and become a full-fledged factor in global competition. Open models of Chinese origin no longer merely meet domestic demand; they are starting to influence which stacks, models and products developers choose far beyond the country's borders.

Support from the top

In the early 2020s, open source became the de facto norm in China's corporate sector. According to a CAICT study, more than 87% of companies used open tools, but only 24% of organizations had dedicated teams to manage open source. The gap was simple: the technology was already widely used, while processes, licensing, security and in-house expertise often failed to keep pace with the scale of adoption.

For Beijing, this became not only an engineering problem but a strategic one. The answer was a state-driven build-out of an ecosystem around open projects. In 2020 the OpenAtom Foundation appeared with the participation of Huawei, Alibaba, Baidu and Tencent, and one of its anchor projects became OpenHarmony, an Android alternative that China sees as a foundation for technological independence.

At the same time, Gitee, the Chinese alternative to GitHub, was growing: by 2020 the platform already had more than 10 million repositories and about 5 million developers. It is far from GitHub's scale, but for the domestic market that was enough to create its own base for development.

From frameworks to models

China did not jump straight to the LLM hype; it came through infrastructure. Even before the wave around DeepSeek, the country had been developing its own deep learning frameworks: PaddlePaddle from Baidu, X-DeepLearning from Alibaba and MindSpore from Huawei. These gave the local market an environment in which large models could be trained and deployed quickly without full dependence on the American stack.

As a result, Chinese teams entered the generative AI era with a ready engineering base and applied use cases in industry, agriculture, cloud and enterprise software.

A few numbers show why this leap got attention worldwide:

• PaddlePaddle is used by more than 23 million developers from 760 thousand companies.
• Training DeepSeek-V3 cost about $5.5 million, according to its authors' estimate, many times less than Western budgets for comparable models.
• A study of downloads of 851 thousand models found that Chinese open-source models accounted for 17.1% of downloads versus 15.8% for American ones.
• By September 2024, downloads of Qwen models had exceeded 600 million worldwide.

The next stage is specialization. Chinese teams are increasingly releasing not only general-purpose chat models but also systems for specific classes of tasks. Examples include the scientific Intern-S1-Pro for chemistry, biology and materials science, the medical Fleming-R1 for diagnosing rare diseases, and Tencent's models for music generation and recognition.

This is an important shift: competition is no longer only about model size or benchmark quality, but about how quickly open-weight solutions turn into working tools for industries.

The West has already plugged in

The main effect is that Chinese open source is no longer an internal story. According to OpenRouter, usage of Chinese open models has grown by about 30% over the past two years from practically zero, with a particularly sharp jump in the second half of 2025. Within a week of its release, DeepSeek burst into the top of chatbot downloads in the US, and some Chinese projects began attracting Western audiences so fast that they had to limit new subscriptions because of a shortage of computing capacity.

Martin Casado, a partner at Andreessen Horowitz, describes the situation this way:

"Startups presenting their open-source projects run on Chinese open models in about 80% of cases."

The behavior of the teams themselves is changing too. Chinese developers used to rarely appear in the English-language public sphere; now they go to Reddit, Western platforms and international communities to promote their models directly. They also monetize successful technical trends quickly: when the OpenClaw agent framework took off in early 2026, Moonshot AI and MiniMax almost immediately offered ready-made solutions requiring minimal setup. This is no longer the stance of a catch-up market, but the behavior of an ecosystem that can quickly productize research advances.

What this means

For China, open AI models have become not a side track but a working mechanism of technological reinforcement. If the trend holds, the global market will face tougher competition not only among closed American platforms but also among open ecosystems, where China already plays the role of one of the centers of power.