From the shock of "DeepSeek" to the Battle of open weight models globally
In 2022, China's artificial intelligence sector faced two consecutive shocks from the United States; the first was the strict export restrictions on advanced electronic chips needed to train large models, and the second was the resounding launch of the ChatGPT robot, which stunned the world and redrawn the boundaries of what is possible in artificial intelligence.
While questions questioned the ability of the Chinese dragon to catch up, in 2024 and 2025, a completely different scene was set, as the world got used to the dominance of Silicon Valley giants such as OpenAI, Google, and anthropic on advanced artificial intelligence models, Chinese open source models – led by DeepSeek and Qwen – rose up to redraw the map of forces, bypassing the idea of competition to the stage of imposing standards.
Does this mean that China has actually surpassed the United States in artificial intelligence models, or are we in front of a new form of leadership that is not measured by performance alone?
First, the boom of open-weight models. From subordination to leadership:
A recent report by the Stanford Institute for Human-Centered Artificial Intelligence (Stanford HAI) revealed that the Chinese system has become a major player in open-weight artificial intelligence models (Open-weight LLMs), and that this poses a real technical and geopolitical challenge for the United States and the world.
After years of relying on meta's Llama architecture as a starting point for the development of its models, Chinese laboratories have moved to a more mature stage, represented in building fully independent models with global competitive capabilities, reflecting an accelerated maturation in research and engineering capabilities.
This transformation culminated in September 2025, when Alibaba's qwen model family surpassed its Llama counterpart to become the most downloaded model globally via the global Huggingface platform.
The data confirms this accelerated growth; Chinese developers accounted for 17.1% of the total global downloads via the (Huggingface) platform during the period from August 2024 to August 2025, surpassing American developers, whose share reached 15.8%, in a clear indication of the shift of the center of gravity of open innovation to the East.
Redefining excellence in artificial intelligence:
The concept of superiority in artificial intelligence has long been associated with traditional measures, which include: the results of standardized tests (Benchmarks), advanced reasoning capabilities, and the dominance of closed superpotential models. But the Stanford Institute report revealed a radical shift in traditional concepts of entrepreneurship.
The report notes that technical sovereignty is now measured by realistic criteria that go beyond theoretical superiority, including: deployment flexibility, cost efficiency, the ability to adapt the model locally, and the momentum it creates in global software communities. Precisely here, Chinese ingenuity emerged, which not only competed, but formulated the criteria for success to suit its vision; it transformed artificial intelligence from a (closed product) to an open-minded (global system), thereby deserving the title of leadership in a new technical reality.
Today, the question in the global artificial intelligence race is no longer: who has the most powerful model behind closed doors It has become: who has the most liberal and widespread model, the ability to reshape the digital structure of the world
Third, Why is China betting on open models?
The report clearly distinguishes two different paths in the development of artificial intelligence models, namely:
1-the path of closed models (Closed Models):
Artificial intelligence in this path will remain a consumer-only product, surrounded by the walls of large companies, which control access and performance, such as: the GPT-4 model from OpenAI, this model cannot access its internal structure, and can only be used via APIs, which imposes strict restrictions on Operation, Control and customization.
2-track of open-weight models (Open-weight Models):
The model in this path turns into a strategic asset that can be fully owned, adapted locally, integrated into sensitive systems with complete flexibility. In this path, the developer of the model gives the developers absolute freedom to deal with it; from loading weights and operating them through a local infrastructure, to modification and retraining without permission.
The genius of the second path lies in the fact that it transforms artificial intelligence from a technical burden and variable cost to a sovereign infrastructure; whoever owns the weights of the model has full control over its data and its digital future, away from market fluctuations or geopolitical pressures.
Fourth, how did China achieve this superiority
The report reveals that Chinese superiority was not born of chance, but came as a result of a smart strategy based on four axes that reshaped the scene, namely:
1-the diversity of companies contributing to the development:
Innovation in artificial intelligence in China is no longer limited to big tech companies, such as (Alibaba) Alibaba, (Baidu) Baidu, but a new front has formed that includes ambitious startups such as (Moonshot AI) and(Z.ai), supported by a huge research momentum from university laboratories. This diversity has created a fertile environment for the production of flexible and highly competitive models.
2-focus on efficiency:
The computing restrictions imposed on China have motivated Chinese developers to focus on developing models that are highly efficient in consuming computing resources, so they relied on the mixture of Experts architecture – known for short as (MoE) – which consumes fewer resources, as it allows only specific parts of the model to be operated during processing.
For example, the flagship qwen3 model has 235 billion parameters, but it activates only 22 billion parameters per input unit, which makes it super-fast and low-cost compared to traditional bulky models.
3-flexible use licenses:
While Western models set complex conditions, China adopts an open doors policy through flexible licenses such as MIT and Apache 2.0. This step removed legal barriers for global companies, allowing Chinese models to be integrated into commercial systems smoothly and without significant restrictions.
