DigitalToday reporter Hyunwoo Choo (추현우) reported that Chinese AI companies are broadening developer ecosystems by promoting open-source models, MIT Technology Review reported on April 21 local time. While U.S. companies maintain an API-based closed strategy, China is offering models whose open-source code can be downloaded, run on users’ own hardware and modified.

The turning point was the R1 reasoning model that DeepSeek released in January 2025. R1 was known to deliver performance comparable to top U.S. models at lower cost, narrowing the gap between Chinese and U.S. AI labs. It also increased developer favorability.

After that, Z.ai, Moonshot, Alibaba’s Qwen and MiniMax followed the same strategy. They are chasing U.S. competitors by releasing higher-performance models one after another. As the AI industry’s focus shifts from pilots to real deployment and integration, the strengths of cheaper tools that are easier to customize are also growing.

Indicators also show this. According to researchers at MIT and Hugging Face, Chinese open-weight models accounted for 17.1 percent of global AI model downloads in the year through August 2025. That narrowly exceeded the U.S. share of 15.86 percent, the first time China led the United States on this metric. Hugging Face data last month showed the number of user-generated derivative models of Alibaba models and the Qwen family was larger than the combined total for Google and Meta models.

There are constraints, of course. Chinese models are influenced by China’s content censorship system and are trained to avoid outputs that run counter to government policy. Anthropic claimed in February that some Chinese labs improperly extracted Claude’s performance through distillation. Distillation is a method that uses one model’s outputs to train another model.

Separate from Western backlash, adoption of Chinese models is increasing in the Global South. AI Singapore, a Singapore government-backed programme, chose Alibaba Qwen instead of Meta’s Llama as the foundation for building the latest regional model. Malaysia said last year it would run its domestic AI ecosystem on a DeepSeek basis. Founders in Nairobi, Sao Paulo and San Francisco are also building services on top of Chinese models.

U.S. companies are maintaining closed strategies, citing the need to recoup massive training costs and concerns about model misuse. By contrast, Chinese companies are drawing in external feedback and contributions more quickly by releasing models as U.S. export controls limit their access to cutting-edge chips. Open-source models are already changing the future of AI to be more multipolar than Silicon Valley expected.