The technological competition between China and the United States has entered a stage in which it is no longer enough to compare which country develops the most sophisticated artificial intelligence models. An increasingly important part of this dispute lies in the way these models are distributed, used, and can be modified by third parties. In this scenario, so-called open-weight models have acquired particular relevance: systems whose trained parameters can be made available to users so that they can download, run, and adapt them, as opposed to closed models whose access remains under the control of their developers. China has turned this strategy into one of the most visible elements of its international expansion in artificial intelligence, with companies such as DeepSeek, Alibaba, and Z.ai developing models that can be used outside their own platforms.
The issue has become more important during 2026 because some Chinese developers are trying to find mechanisms that allow them to preserve the advantages of open weights without ignoring the risks associated with them. In September 2026, Z.ai and Concordia AI presented a specific framework for managing the risks of open-weight models, based on a six-stage process. This proposal is not equivalent to a Chinese law, nor does it mean that China has established a general ban on these systems; rather, it constitutes a risk-management framework developed by actors linked to the artificial intelligence ecosystem. This distinction is important because there are proposals and debates whose scope may still evolve.
What Does It Mean for a Model to Have Open Weights?
To understand why open-weight models are important for China and for international competition, it is necessary to distinguish between an artificial intelligence model and the service through which a company allows it to be used. During the training of a model, millions or billions of parameters are adjusted so that the system learns patterns present in large amounts of information, and these numerical values are commonly referred to as weights. When a company keeps those weights under its control and offers the model exclusively through an interface or an API, it retains considerable ability to apply filters, modify the system, limit functions, or withdraw access. When the weights are published for download, other people can run the model on their own computers or servers without necessarily depending on the infrastructure of the company that developed it.
The expression “open weights” also does not automatically mean that the entire system is open source in the strictest sense. There may be differences between having access to a model’s weights, having full knowledge of the code used to train it, having access to the data used during training, and having detailed documentation about its development. The terms “open source,” “open-weight,” and “open model” can describe different degrees of openness. This distinction is particularly important in the case of China because the strategy of its companies does not depend solely on publishing software, but on allowing developers from different countries to build new applications based on certain models.
Why Is China Betting on This Strategy?
China’s growth in the field of open-weight models cannot be explained solely by a one-off decision by a company. Over the past few years, an ecosystem has developed in China consisting of major technology companies, specialized laboratories, universities, and organizations working on different types of models. DeepSeek became the best-known case internationally following the release of models such as R1, while Alibaba has promoted its Qwen family and other Chinese companies have followed similar strategies. Therefore, talking about Chinese artificial intelligence today does not mean talking about just one company, but rather about a growing group of technology players that compete and seek to expand China’s international presence.
The DeepSeek Effect and the Shift in Perception
The emergence of DeepSeek changed the international discussion because it challenged the idea that having enormous amounts of computing power and the most advanced chips was the only way to develop competitive models. The company attracted considerable attention with models capable of performing reasoning and programming tasks, while using techniques aimed at improving training and inference efficiency. This does not mean that all Chinese models have surpassed all American models, nor that the technological differences between China and the United States have disappeared. What can be observed is that DeepSeek helped demonstrate that competition could also revolve around efficiency, cost, availability, and adaptability.
The subsequent evolution has maintained this trend, although it is advisable to avoid broad conclusions based on a single parameter. During 2026, next-generation Chinese models have emerged from companies such as Z.ai and Moonshot AI, and various technical analyses have placed some Chinese open-weight models close to high-capability American systems on certain tests. Results depend on the benchmarks used, the versions compared, and the specific capabilities being measured, making an absolute ranking between China and the United States difficult to justify based solely on those tests. What is observable, however, is the growth of the Chinese ecosystem and the increasing presence of models developed in China among the alternatives available to developers around the world.
The Paradox of Opening Models and Then Trying to Control Them
Here we encounter one of the main contradictions of this strategy. Publishing the weights allows a model to be studied, adapted, and distributed with much greater freedom, but it also means that the original developer no longer has complete control over what happens after publication. A user can modify the system through fine-tuning, remove certain safety barriers, combine it with other tools, or create a derivative version that is subsequently distributed to other people. In a closed service, a company can modify its filters, suspend an account, or update the model, whereas an open-weight model that has already been downloaded can continue operating even if the original developer changes its policies.
This characteristic means that open models have a different risk profile from systems that operate exclusively on servers controlled by a company. International reports on artificial intelligence safety have indicated that safeguards may be easier to remove from open-weight models and that, once the parameters have been published, the developer loses the ability to effectively retrieve them. This does not mean that all open models are dangerous or that closed models are inherently safe, because both types present different risks. The fundamental difference lies in the developer’s ability to intervene after publication, and it is precisely in this area that China is attempting to develop new management mechanisms.
China’s Six-Stage Proposal
The initiative presented in September 2026 by Z.ai and Concordia AI seeks to address precisely this problem. The framework starts from a simple premise: if weights cannot be withdrawn once they have been published, safety must be assessed before distribution and there must be a specific strategy for risks that may emerge afterward. The proposal organizes management into six stages, ranging from risk identification to governance, adapting procedures that have traditionally been used for models whose access remains under the developer’s control. Its authors present the framework as a basis for balancing the benefits of openness with the problems arising from the loss of control after publication.
The proposal is significant because it recognizes that the safety of an open model cannot depend exclusively on filters applied to an interface. In a closed model, a company can prevent certain requests from reaching the system or block responses before delivering them to the user, whereas a person running the weights locally can modify those layers of protection. For this reason, the proposed approach pays particular attention to identifying potentially dangerous capabilities before publishing the weights, assessing vulnerabilities, and applying mitigation measures. For China, this type of framework may represent an attempt to turn the management of open models into a more systematic process, although it does not yet demonstrate that the problems of control have been resolved.
From Technological Innovation to Regulation
The interest of Chinese authorities in controlling the risks of artificial intelligence did not begin with DeepSeek or with the debate over open weights. In 2023, China approved provisional regulations for generative artificial intelligence services that established obligations related to content, security, data protection, intellectual property, and transparency. These rules apply to certain services offered to the public within China and should not automatically be confused with universal regulation of all existing open-weight models. The regulations nevertheless reflect a constant characteristic of the Chinese approach: promoting the development of artificial intelligence while introducing oversight mechanisms for certain applications.
The Geopolitical Problem: A Technology That Can Be Copied
The international dimension is fundamental because open-weight models do not easily respect traditional economic borders. If a company publishes a model and it achieves widespread distribution, developers in other countries can download and use it without needing to establish a direct commercial relationship with its creator. For China, this represents an opportunity because it allows its technologies to enter foreign ecosystems at a speed that would be more difficult to achieve through services exclusively controlled from China. For the United States and other countries, at the same time, it raises questions about security, technological dependence, privacy, intellectual property, and the possible use of foreign models in sensitive infrastructure.
The issue becomes more complex because open models can have both commercial and scientific and technical applications. A company may use a Chinese model to automate processes, while a researcher may use it to experiment with new artificial intelligence techniques. The possibility of running it locally can reduce costs and provide greater control over data, but it can also make external oversight more difficult. Consequently, the discussion about China and open models is no longer limited to the business sphere, but is connected to technology policy, cybersecurity, and international competition. The distribution of artificial intelligence software can, under certain circumstances, become a form of technological influence that transcends borders.
China and the Competition for Adoption
China’s strategy can also be analyzed from another perspective: the actor that creates the model with the best results across all tests does not necessarily gain the greatest influence, but rather the actor that succeeds in getting its technologies used by a broad community of developers. An inexpensive, accessible, and modifiable model can generate an ecosystem of applications, companies, and derivative projects. Each new project based on that model can contribute to expanding its presence and generating knowledge around its tools. This network effect can be particularly relevant for China because it allows companies from different countries to incorporate Chinese technology into their own products without necessarily having a direct relationship with a Chinese company.
The Risk of Openness Becoming a Disadvantage
The difficulty for China is that the very openness that has helped its models gain users may limit its ability to exercise control over them. If a model is widely distributed, users can modify it for applications that its original creators never anticipated, and some of these changes may reduce or eliminate its safety mechanisms. For this reason, it is reasonable for developers to seek evaluation mechanisms prior to publication, although the question remains as to how much control can actually be maintained after thousands of copies of a model circulate across the Internet. The difficult-to-reverse nature of publishing weights constitutes one of the main technical and regulatory challenges.
This problem is not exclusively Chinese either. The United States, Europe, and other international actors are discussing how to regulate artificial intelligence models capable of producing significant effects on security, the economy, or society, and open-weight models occupy a particular place in that debate. Various government analyses have examined the benefits and risks of making the weights of dual-use foundation models widely available. These studies have pointed out that the availability of weights can promote research and broaden access, but can also make it more difficult to control potential harmful uses once the models are beyond the direct reach of their developers. Therefore, the dilemma facing China forms part of a broader international problem.
Openness Also Has Important Advantages
It would be incomplete to analyze open weights exclusively from the perspective of risk. The ability to download and run a model allows independent researchers to examine its behavior without depending on the limitations imposed by an API, small companies to develop applications without necessarily paying for every query, and organizations to adapt systems to their specific needs. It can also promote competition by reducing dependence on a small group of companies capable of financing enormous computing infrastructures. Under certain circumstances, having a local model can be particularly attractive to organizations that need to maintain greater control over their data.
China vs. the United States: 2 Different Strategies
The comparison between China and the United States cannot be reduced to a simple opposition between open and closed models. The United States also has open-weight initiatives and companies that have experimented with models whose distribution allows different levels of access and modification. Likewise, China has artificial intelligence services that keep certain capabilities under control and do not necessarily distribute all of their components. Even so, there is an observable strategic difference: a significant part of the Chinese industry has turned open-weight models into a tool for technological expansion, while several major U.S. companies have kept some of their most advanced models under controlled access.
China’s decision to pay particular attention to open-weight models reflects a central paradox of contemporary artificial intelligence. Openness can be a tool for reducing costs, accelerating research, increasing adoption, and allowing companies and developers to adapt models to their needs, but that same openness limits the creator’s ability to control what happens after publication. China has achieved a significant international presence through an ecosystem in which DeepSeek, Qwen, Kimi, GLM, and other models have gained visibility, while various analyses show the growth of Chinese models within the open-weight market. However, this growth also forces China to confront an issue that affects all countries: how to manage a technology that, once distributed, can be modified and copied relatively easily.
The initiatives developed during 2026, such as the six-stage framework created by Z.ai and Concordia AI, should be interpreted within this context and not as evidence that China has solved the problem. It is an attempt to build evaluation and risk-management procedures adapted to a technology in which safety measures can be altered after publication and where withdrawing the weights is extremely difficult. At the same time, debates over possible restrictions on the international distribution of advanced models show that China is considering how far openness should go when artificial intelligence also begins to be regarded as a strategic technology. The final outcome of this process has not yet been completely defined, but the direction of the debate is significant: the battle for artificial intelligence is no longer only about developing the model with the best results, but also about deciding who can use it, who can modify it, where it can be distributed, and what controls should exist before and after its publication.
In this sense, open weights could become one of the most relevant pieces of the technological competition between China and the United States. Their value depends not only on the capabilities a model has at the time of its launch, but also on the number of developers who can build upon it and the speed at which those modifications spread around the world. China appears to have identified this characteristic and has used open distribution as a way to expand the international presence of its technologies, while the United States and other countries are studying how to respond to the risks and opportunities generated by this new form of development. The decisive question in the coming years will not only be which country develops the most advanced models, but also which combination of openness, security, responsibility, and regulation will make it possible to take advantage of their benefits without ignoring the problems that arise when an advanced technology is no longer under the exclusive control of those who created it.
For companies seeking to understand how these changes may affect their processes, infrastructure, and technology strategy, having specialized support is becoming increasingly relevant. ITD Consulting provides technology services and IT solutions aimed at organizations that need to evaluate new tools, optimize their infrastructure, and take advantage of opportunities associated with digital transformation. In an international context marked by the rapid advancement of artificial intelligence and technological competition between China, the United States, and other markets, properly analyzing opportunities and risks can help companies make more informed technology decisions. To learn more about ITD Consulting’s services and discuss your organization’s specific needs, you can write to [email protected].
EN 
ES