This year, China’s open source large models have been intensively iterated: starting with DeepSeek/R1 in January, followed by the successive releases of Alibaba Qwen, Moonshot, Z.ai, MiniMax, etc. The above models generally provide open source/open-weighted versions that can be downloaded and modified for free, promoting the development of global open source models. In the United States, manufacturers that have long adhered to a closed-source strategy are feeling pressure, and OpenAI responded by launching its first open source model, gpt-oss, in August.
The history of science and technology shows that the battle for standards is not always decided by "the strongest technology", but availability and flexibility are often the key. This is why China’s advancement in open source AI has attracted the attention of U.S. government and enterprises. The U.S. AI Action Plan released in July also proposed that open source models are expected to become global standards in some business and academic scenarios, and called on the United States to create a so-called "open source model based on American values."
In terms of business model, open source winners have limited short-term direct returns, but they can monetize supporting services (applications, cloud and tool chains) through the free part, just like the Android/Linux ecosystem. The research community has long embraced open source to accelerate the evolution of cutting-edge technologies; China also encourages open source R&D in fields such as AI, operating systems, semiconductor architecture, and engineering software.
Enterprise-side adoption is expanding. Overseas Chinese Bank of Singapore (OCBC) has self-developed about 30 internal tools based on open source models: Gemma for document summarization, Qwen for assisting in code writing, and DeepSeek for market analysis. Its strategy is to avoid being locked into a single model, switch to new versions at any time, and prefer models that are familiar to most developers and easy to obtain technical support.
In terms of performance experience, the third-party evaluation organization Artificial Analysis pointed out that since November last year, in the comprehensive scores of authoritative evaluations, the overall running scores of China's large models with public weights are higher than the current strongest open source models in the United States; in one of its comparisons, the Qwen3 version is better than gpt-oss. Many Asian engineers also reported that in the Chinese and regional cultural context, some Chinese models have a better grasp of implicit intentions and polite expressions. When Shinichi Usami, an engineer in Yokohama, Japan, built a customer service robot for retail customers, he chose Qwen because of its better understanding of delicate expressions.
Industrial competition is extending from closed source price wars to competition for users in the open source ecosystem. Researchers say that Chinese companies often prioritize increasing user stickiness and then undertake commercialization with supporting services. During the window period, start-ups have the opportunity to quickly accumulate users, but large platforms have greater liquidity. Andrew Ng, head of Silicon Valley startup DeepLearning.AI, argued in a recent blog that fierce competition will eliminate a group of players, but will also forge stronger companies.