ChatGPT developer OpenAI has entered the open-source AI ecosystem traditionally led by Chinese AI corporations by releasing two types of AI models based on 'open weight'. This marks OpenAI's first foray into an open model in six years, following its previous insistence on paid closed models. Amid escalating U.S.-China AI hegemony competition, this move can be interpreted as a response to the Trump administration's AI promotion policies and a strategy to keep pace with Chinese AI corporations, led by DeepSeek, which are rapidly releasing high-performance open-source AI models.
◇ OpenAI releases 'gpt-oss' series… "the best performing among open weight models"
According to the AI industry on the 7th, OpenAI released the first partially open inference model, the 'gpt-oss' series, the previous day. The 'gpt-oss-120b' and 'gpt-oss-20b' released by OpenAI through the AI software sharing platform Hugging Face are 'open weight' models that deliver powerful performance at relatively low expense.
Existing major AI models from OpenAI, such as GPT-3, GPT-4o, and o3, operate within a closed system where model parameters, weights, training data, and source code are not shared at all. The newly introduced open weight model is not completely open source, but it is classified as partially open since weights are publicly available. Developers, researchers, and corporations can download the model weights (the core values learned by the model) and modify them for custom use.
Sam Altman, CEO of OpenAI, stated on social media platform X (formerly Twitter) that "'gpt-oss' is the most advanced open weight inference model with the best performance and versatility in the world." He emphasized that the performance of the 'gpt-oss' series is comparable to that of one of its flagship models, 'o4-mini', and that it is relatively compact and efficient enough to run on high-spec laptops.
◇ China expands influence with open-source AI… Trump administration takes notice
OpenAI, which has focused on monetizing its paid closed models until now, has shifted toward a 'partially open' direction, influenced significantly by the U.S. government's AI strategy and the early year shock from DeepSeek.
The AI action plan, recently announced by President Donald Trump, prioritizes the 'dissemination of open-source and open weight AI.' This is based on the premise that the U.S. must establish an open AI model ecosystem to win the AI hegemony battle against China. Specifically, the strategy involves encouraging individuals, researchers, and corporations within the U.S. to freely utilize open-source AI models to promote private-sector innovation, while exporting closed AI models abroad to curb China and maintain U.S. AI dominance.
The U.S. government emphasized in the AI action plan documents that "the United States must possess a leading open-source model based on American values," noting that "open-source and open weight models can become global standards in certain industries and worldwide academic research, thus possessing geostrategic value."
In particular, the 'DeepSeek shock' that swept the global AI industry at the beginning of the year is identified as the main reason driving the U.S. government's push for an open-source AI dissemination strategy. After DeepSeek released the low-cost, high-performance open-source inference model 'R1', American researchers and developers actively utilized it, leading to R1 becoming the most popular model on Hugging Face, with thousands of variant models subsequently used by major corporations and research institutions, heightening fears that the U.S. could lose its AI leadership to China. Tech publication VentureBeat described the situation at the time as "the moment American AI was built for the first time on a Chinese model."
China is expanding its influence in the global market by promoting open-source AI. The Washington Post cited analysis from the global AI research organization 'Artificial Analysis', reporting that "of the top 15 AI models currently, only 5 are open-source, all developed by Chinese corporations." In fact, in July alone, Chinese corporations like Alibaba and Baidu unveiled four new open-source models, while major U.S. tech companies, excluding Meta, continue to maintain closed strategies.
◇ OpenAI shifts to a mixed strategy of closed and open-source to maintain market advantage
There is also an analysis that OpenAI has adopted a mixed strategy of providing both closed and open-source models to maintain an advantage in the AI ecosystem, recognizing the limitations of its closed strategy. After launching ChatGPT in 2022, OpenAI rose as a dominant force in the closed AI ecosystem, but recently competitors such as Google's Gemini, Anthropic's Claude, and xAI's Grok are rapidly closing the gap, while in the open-source AI market, Chinese corporations led by DeepSeek are making significant advances, threatening OpenAI's previously solid industry leadership.
Following the DeepSeek incident, CEO Altman acknowledged, "We thought we were on the wrong side of history" before embarking on the development of open-source models. While closed models, which exclusively manage the structure and internal workings of AI models, are advantageous for revenue generation, open-source models that disclose source codes and designs allow developers to utilize them freely, aiding in technology diffusion and user acquisition.
Meta, the parent company of Facebook, is also shifting towards a mixed strategy. While Meta has sought differentiation with its open-source model 'LLaMA', it recently indicated the potential introduction of closed models for the first time. Mark Zuckerberg, CEO of Meta, is committed to developing 'superintelligent AI' that surpasses human intelligence, mentioning that related technologies may not be disclosed as open-source. The astronomical costs invested in superintelligent AI are utilized for development under a closed model for commercialization and revenue generation, while LLaMA is provided as open-source to acquire more users, executing a 'two-track' strategy.
CEO Zuckerberg stated, "The benefits of superintelligence should be widely distributed, but at the same time could create new risks," adding, "We will carefully decide which models to disclose as open-source."