Chinese government and AI companies have increasingly embraced openness as a key strategy amid the global AI competition. However, due to commercial pressure and associated risks with open-weight models, Beijing is likely to pursue a hybrid and tiered ecosystem where open and closed models coexist. This would allow China to broaden international adoption without giving up the commercial and strategic value of their most advanced models.
In July, the 2026 World Artificial Intelligence Conference (WAIC) opened on the banks of the Huangpu River in Shanghai. In his keynote speech, Chinese President Xi Jinping said that AI development should not be “a solo performance by a single country, but a symphony of international cooperation”, and urged countries to embrace “open source, openness, collaboration and sharing” in AI. His remarks once again brought the debate over AI openness and the future of open-weight models into the international spotlight.[1]
Chinese AI companies have increasingly embraced openness as a key part of their strategies, and in some areas have emerged as global leaders in open-weight AI. On 16 July 2026, Beijing-based startup Moonshot AI released Kimi K3, a 2.8 trillion-parameter model that it claimed to be the world’s largest open-weight AI system. At the time of writing, ranking data on the platform OpenRouter showed that the five most-used open AI models worldwide were all Chinese models.
During our recent research visits in the UK, policy analysts repeatedly asked us the same question: Will Chinese language models remain open-weight in the long run? To answer this question, we look at how the Chinese government views open-weight models, discuss what China (and its AI firms) gains and risks by keeping model weights open or closed, and argue that Beijing is likely to pursue a hybrid and tiered strategy rather than commit fully to either path.
The Chinese Government’s Stance on Open-weight AI Models
China has consistently supported open-weight AI as a way to encourage innovation, strengthen the AI ecosystem, and expand international cooperation.
Domestically, openness is treated as a way to broaden access to AI and accelerate its adoption and diffusion. In 2024, the Ministry of Science and Technology encouraged firms to open up industry-specific data, models and algorithms, and called on leading AI companies to help build more open AI ecosystems. Later that year, the Ministry of Industry and Information Technology, Ministry of Finance, People’s Bank of China and National Financial Regulatory Administration jointly issued an action plan that called on small and medium-sized ewjnterprises to participate actively in open-weight projects and aimed to lower barriers to AI development and deployment. This emphasis continued in the 2026 “AI+Manufacturing” initiative, which set the goal of building a world-leading open-source ecosystem by 2027.
Internationally, the government increasingly presents openness as part of China’s broader agenda of global AI cooperation. At the 2025 WAIC, Premier Li Qiang said that “China would be more open in sharing open-source technologies and products”. The subsequent State Council guideline on the “AI+” initiative linked accessible AI technologies to international cooperation on computing power, data and talent, with particular emphasis on capacity-building in the Global South. The Chair’s Statement of the 2026 WAIC went further, describing openness as an important pathway towards inclusive AI development and calling for more open and internationally connected AI communities.
However, official language does not imply that every Chinese model should be open-weight, as it also emphasizes intellectual-property protection, security, responsible governance and companies’ freedom to choose their own development strategies. Taken together, Beijing appears to treat model openness as a strategic choice rather than an across-the-board principle, leaving room for both open-weight and closed models.
The Commercial Pressure and Risks of the Open-weight Strategy
That flexibility on openness may become important as keeping models open-weight grows more costly, with US-China competition pushing frontier models toward the trillion-parameter scale and the spread of AI agents sharply increasing model use.
The commercial strain has already become visible. Opening a model’s weight can expand its user base, attract developers and strengthen an ecosystem. Yet it is generally harder to capture recurring revenue from an open-weight model than from a proprietary service built around subscriptions and API fees. Moonshot illustrates the tension. Just three days after Kimi K3’s launch, the company temporarily paused new subscriptions as demand strained its available computing capacity, while prioritizing existing subscribers. DeepSeek has faced similar capacity constraints during periods of surging demand, at one point suspending API top-ups because of server-resource pressure.
When an official platform cannot absorb all of that demand, users can move to third-party deployments and service providers. The original model developer then bears much of the cost of R&D and ecosystem-building while some users and some of the commercial value they generate shift elsewhere. Wider availability also intensifies price competition, putting further pressure on margins.
Chinese companies are therefore beginning to experiment with ways of keeping models open-weight while trying to retain more of their commercial value. The Kimi K3 licence is one example. Commercial products or services built on the model must display the “Kimi K3” prominently once they exceed either 100 million monthly active users or US$20 million in monthly revenue. A separate provision applies to licensees or affiliates operating a Model-as-a-Service business: if aggregate revenue from that business exceeds US$20 million over any consecutive twelve-month period, a separate agreement with Moonshot is required before commercial use. Reuters reported on 7 August that Alibaba was planning a comparable revenue-sharing arrangement for major commercial users of its next open model, Qwen3.8-Max.10.
These moves suggest that “open-weight” no longer has to mean giving away most of the commercial upside. Chinese model developers are beginning to search for a new balance between broad ecosystem adoption and sustainable returns.
The second pressure concerns the allocation of rights and responsibilities. Once model weights have been released, however, access cannot simply be withdrawn in the same way that an online service can be switched off. Third parties can download the model, fine-tune it, alter its safeguards and deploy it in environments over which the original developer has little or no control.
Open-weight models are also more exposed to technical scrutiny: the availability of model weights makes it easier to investigate training data or identify signs that a model may have been distilled from another system. These practices can more readily spark disputes over model provenance, permissible reuse, ownership, and responsibility, with such issues becoming increasingly salient amid intensifying great-power AI competition.
Open-Weight as a Source of Competitive and Strategic Advantage
However, China is unlikely to follow the US model of moving broadly towards closed-weight AI.
One reason is that efficiency has become one of the clearest competitive advantages of China’s AI industry. Much of the strength of Chinese technology companies lies in compute efficiency, training engineering, inference cost, model compression, and software-hardware adaptation. This produces a different industrial logic from that of leading US frontier labs. While American companies such as OpenAI and Anthropic compete at the cutting edge of compute, advanced chips and proprietary API services, many Chinese developers have stronger incentives to differentiate themselves through lower cost, local deployability and the ability to adapt or further develop a model downstream.
This has made Chinese models particularly appealing in markets where concerns over data sovereignty and dependence on foreign technology are especially pronounced, including much of the Global South. Open or locally deployable models can reduce dependence on a single foreign provider, give users more control over where data is processed, and make it easier to adapt systems to local languages and use cases. These advantages have become an important route through which Chinese models attract international developers and users. For that reason alone, Chinese firms are unlikely to give up openness lightly.
There is also a broader political and diplomatic logic. China’s effort to build an open AI ecosystem fits closely with its blueprint on global governance, from the idea of a community with a shared future for humanity and the BRI to the principles of consultation, joint contribution and shared benefit, as well as the Global Development Initiative, Global Security Initiative, Global Civilization Initiative and Global Governance Initiative.
China describes itself as a provider of international public goods relating to AI. The future trajectory of China’s AI models therefore cannot be read through technology and commercial incentives alone. Openness is also part of a wider national strategy: it helps expand developer ecosystems and international adoption while reinforcing China’s preferred narrative of more inclusive access to AI.
From Open Weights to Real Access
Returning to the question we began with, the answer is unlikely to be a simple “yes” or “no” when it comes to model openness. Rather, China is likely to move towards a hybrid ecosystem in which open and closed models coexist. Small and mid-sized foundation models, many industry-specific models, and models designed to attract developers and expand application ecosystems will probably remain relatively open. By contrast, the most advanced models – those requiring the greatest R&D investment, posing the highest safety risks, or offering the strongest prospects for monetization – are more likely to be closed, released with a delay, distributed under restrictive licenses, or made available only via APIs. ByteDance already offers an early example of this tiered approach. It has released the weights of selected models such as Seed-OSS, while keeping its flagship models proprietary and providing access through consumer products such as Doubao and API services via Volcano Engine.
Such a tiered approach would not conflict with China’s broader commitment to making AI more widely accessible. For many countries, especially in the Global South, opening the weights of models with hundreds of billions or even trillions of parameters will offer limited practical benefit if they lack the computing power, data-center capacity and engineering resources needed to run them at scale. What matters more is access to capable models that can operate under tighter constraints on compute, electricity and network capacity, while still being adapted to local languages and applications.
This helps explain why Chinese AI companies are placing growing emphasis on model efficiency and what is called “intelligence density” (智能密度): the capability delivered for a given model size or level of compute. Based on the Intelligence Index vs. Cost per Intelligence Index Task test given by artificialanalysis.ai, DeepSeek-V4-Flash significantly outperforms all other models in terms of intelligence density. High intelligence density made DeepSeek-V4-Flash the most popular model on OpenRouter, highlighting the global demand for high intelligence density models.
From this perspective, China may have stronger incentives to keep highly capable, efficient mid-sized models open-weight while keeping the most expensive frontier models closed or more tightly controlled. This would allow Chinese firms to broaden international adoption without giving up the commercial returns and strategic value of their most advanced models.
[1] In this article, we primarily use the term “open-weight”refers to models whose trained parameters (or “weights”) are publicly released and available for users to download and run. This does not mean that the model’s training data, training code, or full technical documentation is also publicly available. The term is therefore narrower than “open-source AI.”