From Policy Blueprint to Industrial Practice: China's Embrace of Open-Weight AI

Introduction

In 2017, China's New Generation Artificial Intelligence Development Plan enshrined the rapid development of Artificial Intelligence as a strategic priority. It laid out a blueprint for how PRC authorities planned to deploy industrial policy tools to build a first-mover advantage in a technology set to reshape economic development, social construction, and international competition. By 2030, China should be the world's primary innovation centre, leading in AI theories, technologies, and applications. 

What has been particularly notable in China's policy approach is its emphasis on open-source AI: the 2017 plan makes 'open-source sharing' a foundational principle, targeting collaborative development ecosystems that would facilitate technological innovation. The development of open-source AI is featured in the 14th Five-Year Plan, and in recent high-level AI policies such as the 2024 AI Plus Initiative and its implementation guideline. On the international level, open-source AI has formed a pillar of China's approach to global AI governance.

Such longstanding government support now appears to increasingly match industrial reality. Spurred in part by the release of DeepSeek's V3 and R1 models in December 2024 and January 2025, China's corporate AI ecosystem has undergone a shift toward open-source releases. Tracing this development to both business and government strategy reveals a complex interplay of driving forces that have increasingly consolidated the centrality of open-source approaches in China's AI development. These rationales, means, and limits of how the PRC and industry players have embraced open-source AI have crucial implications for the future of AI industrial development and global governance. 

Definitions: Open-Source, Open-Weight, and Closed-Source AI Systems

According to the Open Source Initiative, an AI system qualifies as open source if it provides users with four freedoms: to use the AI system for any purpose, to study its inner workings, modify or adapt it to specific needs, and to share it for the use of others. Open-source AI systems require the release of sufficient information about training data, source code, and parameters – internal variables acquired during the training process that shape how models interpret data, such as weights and biases. They are provided under open-source licences with permissive terms for use, modification, and distribution. Together, this allows end users to run open-source AI models locally and modify parameters to suit specific use cases. Open-source systems offer high transparency and affordable pricing, facilitating community contributions and research use. 

By contrast, closed-source or proprietary AI refers to models whose data sets, algorithms, and weights are not freely accessible. Examples include the mainline models of U.S. developers like OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini. Developers monetize their proprietary models through applications and designated Application Performance Interfaces (APIs), with users generally paying a subscription or per-token prices. 

In practice, however, AI systems exist on a 'gradient of system access'. Most models labelled as open-source are, in fact, more accurately categorised as open-weight: they publicly share model weights but do not release their training data and source code, or impose certain use restrictions under their respective licences. While Chinese models are commonly described as open-source, all existing releases have, strictly speaking, been open-weight. Regarding the PRC, we can thus distinguish between political and industrial commitments to open-source approaches on the one hand, and specific open-weight AI models on the other. Open-sourcing, in turn, refers to the public release of formerly closed-source models – a term which can be applied to a range of open release frameworks. Open-sourcing strategies adopted by Chinese AI developers have generally involved the public release of model weights and source code, partially or in its entirety, under permissive licences, without disclosing the underlying training data.

Outside of the PRC, closed-source approaches continue to dominate. Most developers have either focused exclusively on proprietary releases or adopted a hybrid approach with closed-source and open-source releases, often of lower-performing models. Among U.S. companies, Meta has been a front-runner in open-weight AI with its Llama family of models. Other examples include Google's Gemma line of lightweight models and OpenAI's recent gpt-oss family. Outside the U.S., French company Mistral AI has released several open-weight LLMs. 

Mapping the Evolving Role of Open-Weight AI in China's Industrial Ecosystem

China's AI landscape comprises a diverse set of players – from established tech giants to startups. This industrial ecosystem is currently experiencing a shift towards open-weight model releases, kick-started by the unveiling of DeepSeek's V3 and R1 models. While not the first open LLMs to enter the Chinese market, DeepSeek-R1's performance, uptake, and disruptive pricing have upended established company strategies: Baidu and Moonshot AI have each abandoned their commitment to a closed-source approach, while companies like Alibaba are open-sourcing increasingly high-performing models. The vast majority of Chinese AI players now employ hybrid strategies that combine proprietary offerings with a growing suite of open-weight releases.

China’s Big Tech companies – most prominently Alibaba, Tencent, Baidu, Huawei, and ByteDance – have each invested heavily in advancing their AI capabilities. Among them, Alibaba Cloud (Aliyun) has been the most prominent proponent of open-source AI development. Having shifted to a hybrid approach for its Qwen2 and Qwen2.5 models, Alibaba's recent releases indicate a trend towards open-sourcing increasingly high-performing models. While the company maintains proprietary offerings, its recent Qwen3 series, unveiled in April, is built on an open-weight framework. With its main business line in cloud computing, Alibaba Cloud offers AI inference on its native cloud hosting platform. Tencent debuted its Hunyuan series of foundation models in September 2023. While traditionally focused on offering proprietary access through Tencent Cloud, the company has open-sourced several high-performing models in recent months, ranging from translation to 3D generation. China's internet search giant Baidu was an early frontrunner in AI development, but has since struggled to keep pace with competitors. Notable for its commitment to closed-source AI development, the company recently announced it would shift to an open-source approach for its next-generation Ernie models, citing the success of DeepSeek in driving adoption. Finally, Huawei and ByteDance – both traditionally focused on proprietary development – respectively open-sourced [1][2] selected models in the past months. 

The emergence of China's AI startup scene has been dominated by the so-called 'AI tigers' – MiniMax, Baichuan, Zhipu AI, and Moonshot AI – with valuations exceeding $1 billion. This status quo was upended in January 2025 when Hangzhou-based startup DeepSeek achieved international recognition following the release of its open LLMs DeepSeek-V3 and DeepSeek-R1. Founded in 2023, DeepSeek has focused exclusively on open-source model development, diverging from the hybrid strategy commonly adopted by most Chinese companies. It has placed a particular emphasis on model efficiency to reduce training costs and compute requirements, allowing it to offer highly competitive API pricing. China's incumbent unicorns have responded variably to these competitive pressures by open-sourcing and shifting their focus to the application side. MiniMax and Moonshot AI, which traditionally followed a closed-source approach, have recently committed[1][2] to gradually open-sourcing their model development. Both have cited[1][2] benefits for innovation, research input, and application development. Beijing-based Zhipu (Z.ai) similarly announced in June 2025 that it would open-source its GLM-4.5 model, while maintaining a suite of proprietary offerings. Baichuan has a history of open-source development, having released China's first open-weight LLM in June 2023. The company has since shifted its focus from model development to enterprise applications, mirroring the pivot announced by fellow AI start-up 01.AI in January 2025 and a growing focus on agentic AI applications across the industry.

Converging Government Policy and Industry Incentives

An open-source approach to AI development offers several potential benefits for enhancing the speed of innovation and deployment. Open-source ecosystems allow model developers to build upon each other's technological advances, creating competitive pressures that push companies towards faster iterations and innovative model architectures. Open releases also enable companies to draw upon insights from a broader, international community of developers, who independently develop fixes or derivative model versions. Beyond the development stage, open-source and open-weight models facilitate adoption and integration into applications. For enterprise clients, adaptability is a key factor that enables them to tailor models to their specific needs and computing resources. Compared to closed systems, open-weight models' adaptability, relative transparency, and pricing make them more suitable for research use. When it comes to broader societal adoption, the lower cost of APIs facilitates the development of applications built on top of AI models or their integration, e.g., into consumer electronics. 

These characteristics align with Chinese government priorities along multiple dimensions: for one, China's support for open-source AI mirrors its broader embrace of open-source software and technology as a means to reduce dependence on foreign technology companies and proprietary technologies, which may be subject to export restrictions. The PRC's approach thus leverages the benefits of open-source technologies to foster collaboration and innovation, thereby accelerating technological progress towards a self-sufficient domestic ecosystem. The existing integration of Chinese developers and technological players in global open-source software communities, in turn, has facilitated the adoption of an open-source approach to AI. 

Below this overarching goal, recent Chinese policy has emphasised AI adoption. According to the August 2025 AI Plus blueprint, China seeks to leverage AI as a tool for economic and industrial transformation, foster scientific innovation, and drive consumption through AI-powered consumer products and services. Each of these areas requires broad deployment of AI models across research, enterprise users, and consumers. Adoption, in turn, has an international dimension: open-weight models can achieve a wider diffusion on the international level, with lower-cost pricing opening use cases particularly in the Global South. China's embrace of open-source AI thus strengthens its credibility in advocating inclusive approaches to global AI governance, while enhancing its de facto influence on the development of technical AI standards. 

Open-source AI development thus emerges as a potential instrument to promote China's strategic policy aims centred on driving innovation and adoption. From the business side, the case for open-source development is less clear-cut. While it complicates monetisation and the protection of proprietary technology, there are also substantial arguments for companies to adopt open-weight offerings, particularly as part of a hybrid strategy. Their increasing salience may help explain the flurry of open-weight releases across China's AI ecosystem in recent months.

Open models present specific challenges from a business standpoint: with users free to run models on local computing hardware, APIs for open-source and open-weight models are generally priced at a very low per-token cost. This makes it difficult to achieve initial monetisation, which is needed to recoup investments in AI training. Increased transparency regarding parameters and technical documentation – while driving collaboration – also allows competitors to build upon in-house developments, running against the imperative to protect proprietary technologies. However, two factors relativise this assessment regarding the business case for open-source AI: the distinct benefits of open-source approaches for technological innovation and market adoption, combined with the fact that open-source releases are often not standalone but form part of hybrid strategies or a longer-term business model. 

Many of the aforementioned benefits of open-source approaches are equally applicable at the company level. Open-source and open-weight models invite contributions, fixes, and other post-release feedback from a broader community of developers, allowing innovation to progress at scale and speed beyond the capacity of in-house development teams. Combined with adaptability and lower-cost pricing, open releases thus facilitate adoption and incentivise third-party developers to create applications on top of a given model. For AI companies, drawing users into publicly available open-weight solutions has two primary benefits with a view to monetisation: open releases build reputation and demonstrate technical capabilities. They may also contribute to developer lock-in, as developers who build applications around certain AI models become integrated into a specific company's model ecosystem. 

This, in turn, links to the role that open-weight AI offerings may play in hybrid strategies or longer-term business models. Open-weight AI releases can serve as pathways for monetisation. By building trust and integration with users, they may drive business to a company's proprietary models or ancillary services, including private AI deployments, fine-tuning, and other enterprise tools. In the longer-term, lock-in through application ecosystems built around a company's AI models may facilitate monetising access to these models and to the services structured around them in the future. Finally, AI models can also draw clients to parallel lines of business. As one of China's leading promoters of open-source AI, Alibaba Cloud has leveraged open-weight releases to encourage developers to adopt its models, thereby driving traffic to its cloud-hosting services.

Conclusion and Outlook

The PRC has adopted an open-source approach to AI development as part of its industrial policy toolkit, aiming to promote rapid technological innovation and drive the flow of AI into consumer and industrial applications. Open-weight models are instrumental in China's blueprint for achieving technology leadership in AI, leveraging the technology to support its economic transformation objectives and secure global influence in AI governance and standard-setting. 

Despite such high-level political backing, closed-source models have remained important, particularly in the early years of the industry. An overview of current industry dynamics reveals a broad rebalancing towards open-weight releases, spurred by the January 2025 release of DeepSeek's R1 model. Among AI start-ups and Big Tech companies alike, hybrid strategies dominate, with a growing trend towards open-sourcing increasingly frontier models. 

Tracing the political and business case for open-weight AI reveals intertwined yet disparate driving forces that have increasingly coalesced to cement the role of open models in China's AI industry. With open models offering benefits for innovation and adoption as well as competitive challenges, China's tech companies have each struck specific balances between openness and protection. With the release of DeepSeek-R1 demonstrating how open-weight systems can achieve high performance and drive uptake, and exerting pressure on incumbent industry strategies through disruptive pricing, these established balances have shifted: China's AI companies have increasingly integrated open-weight releases into their business cases, driving traffic to proprietary solutions and lines of business or seeking to cement industry positions through user uptake and application development. 

While the PRC's embrace of open-source development has allowed it to drive rapid innovation and narrow the gap with leading U.S. models, internal and external stumbling blocks persist. For one, concerns about data privacy or regulatory restrictions may limit the adoption of Chinese APIs and AI-powered products by international users. Following its release in January, several countries quickly restricted the use of DeepSeek's AI App on government networks. On a larger scale, such barriers may challenge China's ability to build influence through open-source releases. However, the most crucial risk factor is how PRC policy itself will evolve: the ambivalence of China's stance towards decentralised global development communities is evidenced by efforts to develop domestic alternatives to popular open-source platforms such as GitHub and Hugging Face, the latter of which was blocked in China in 2023. With open-weight systems facilitating propagation and user modification, they are ill at ease with government efforts to more closely regulate generative AI models. The question remains how authorities will approach state control, censorship, and liability for open-weight models without stifling their growing role in China's tech ecosystem. 

This article does not necessarily reflect the opinions of European Guanxi, its leadership, members, partners, or stakeholders, nor of those of its editors or staff. They have been formulated by the author in their full capacity, and shall not be used for any other purposes other than those they are intended for. European Guanxi assumes no liability or responsibility deriving from the improper use of the contents of this report. Any false facts, errors, and controversial opinions contained in the articles are proper and exclusive of the authors. European Guanxi or its staff and collaborators cannot be held responsible or legally liable for the use of any and all information contained in this document.

ABOUT THE AUTHOR

Franziska Kändler is a graduate of Sciences Po Paris, specialising in digital and technology policy with a particular focus on EU-China and U.S.-China relations. She has worked at the European Space Policy Institute and the Prague Security Studies Institute, and is the U.S.-China Analyst of the Heidelberg Institute for International Conflict Research. 

This article was edited by Robin Millet and Nika Tatoshvili.

Featured Image: Emmanuel R. Goffi, keynote at the World AI Conference (WAIC) 2025 in Shanghai / Free for Use / Creative Commons Attribution-Share Alike 4.0 International license