This is the third and final blog in a three-part series on the historical progress of China’s open source community since the “DeepSeek Moment” in January 2025. The first blog on strategic changes and open artifact growth is available here, and the second blog on architectural and hardware changes is available here.
In this third article, we examine the trajectory and trajectory of China’s prominent AI organizations and suggest future directions for open source.
For AI researchers and developers who contribute to and rely on the open source ecosystem, and for policymakers who understand the rapidly changing environment within their organizations and the interests of the global community, open source will be the dominant and popular approach for Chinese AI organizations in the near future. Openly sharing artifacts, from models to papers to deployment infrastructure, creates strategies targeted for large-scale deployment and integration.
China’s organic open source AI ecosystem
Our exploration of DeepSeek’s strategic and architectural changes since R1 provides our first glimpse into how an organic open source AI ecosystem is taking shape in China. The culmination of strong players, including established players in open source, new players, and players who change tack entirely to contribute to a new open culture, shows that an open collaborative approach is mutually beneficial.
This collaboration extends beyond borders. The most followed organization on Hugging Face is DeepSeek, and the fourth most followed organization is Qwen.
In addition to models, openly sharing science and technology informed not only other AI organizations but also the open source community at large. The most popular papers on Hugging Face primarily come from Chinese organizations such as ByteDance, DeepSeek, Tencent, and Qwen.
Source: https://huggingface.co/spaces/evijit/PaperVerse
established
Alibaba has positioned open source as an ecosystem and infrastructure strategy. Qwen was not formed as a single flagship model, but was continually expanded into a family covering multiple sizes, tasks, and modalities with frequent updates with Hugging Face and our proprietary platform ModelScope. Its influence was not concentrated in a single version. Instead, it was repeatedly reused as a component across a variety of scenarios, gradually taking on the role of a common AI foundation. By mid-2025, Qwen will be the most derived model on Hugging Face, with over 113,000 models using Qwen as a base and over 200,000 model repositories tagged to Qwen, far exceeding Meta’s Llama’s 27,000 and DeepSeek’s 6,000. Alibaba boasts almost as many derivatives across its organization as Google and Meta combined.
At the same time, Alibaba coordinated model development with cloud and hardware infrastructure, integrating models, chips, platforms, and applications into a single engineering stack.
Tencent has also made a major shift from borrowing to building. Tencent, one of the first major companies to integrate DeepSeek into a major consumer product after the release of R1, did not initially frame open source as a public narrative. Instead, we introduced a mature model through plugin-style integrations, performed extensive internal validation, and only then began releasing our own features. Since May 2025, Tencent has used its own brand named Tencent Hunyuan (now Tencent HY) to accelerate open releases in areas where it already had strengths, such as vision, video, and 3D, and these models were quickly adopted by the community.
Following an “AI application factory” approach, ByteDance has begun to selectively open source high-value components while maintaining a competitive focus on product entry points and large-scale use. In this context, the ByteDance Seed team has contributed several notable open source artifacts, including UI-TARS-1.5 for multimodal UI understanding, **Seed-Coder** for data-centric code modeling, and the SuperGPQA dataset for systematic inference evaluation. Despite its relatively low profile open source presence, ByteDance has achieved significant scale in China’s AI market, with its AI application Doubao exceeding 100 million DAU in December 2025.
The most notable change also began within Baidu, where the CEO openly expressed his disdain for open source. After years of prioritizing a closed model, Baidu has re-entered the ecosystem through free access and open releases such as the Ernie 4.5 series. This transition was accompanied by new investments in PaddlePaddle, an open source framework, and Kunlunxin, a proprietary AI chip that announced its IPO on January 1, 2026. By connecting models, chips, and PaddlePaddle in a more open system, Baidu can lower costs, attract developers, and influence standards while maintaining strategic control under common constraints of compute, cost, and regulation.
The normality of a “deep seek moment”
Among the startups, Moonshot, Z.ai, and MiniMax adapted quickly and brought new momentum to the open source community within months after R1. Models like Kimi K2, GLM-4.5, and MiniMax M2 all made it to AI-World’s Open Model Milestone rankings. At the end of 2025, Z.ai and MiniMax released their most advanced open source model to date, and then announced IPO plans in quick succession.
Kim K2’s open sourcing was widely described as a “new DeepSeek moment” for the community. Moonshot has not announced an IPO, but market reports indicate the company will raise approximately $500 million by the end of 2025, with AGI and agent-based systems positioned as key commercialization targets.
Application-first companies such as Xiaohongshu, Bilibili, Xiaomi, and Meituan, which previously focused only on the application layer, have started training and releasing their own models. The availability of powerful inference at low cost through open source has made building in-house models practical due to its native advantages in real-world usage scenarios and domain data. Tailor AI to your specific business rather than being constrained by the cost structure or limitations of an external provider.
As the business world seizes the opportunity for positive ROI growth, research institutions and the broader community will welcome the change more enthusiastically. Organizations such as BAAI and Shanghai AI Lab have focused more efforts on toolchains, evaluation systems, data platforms, deployment infrastructure, and projects such as FlagOpen, OpenDataLab, and OpenCompass. These efforts were not about the performance of a single model, but about strengthening the long-term foundations of the ecosystem.
Foundation for the future
The new ecosystem is not characterized by an increase in models, but by the formation of entire chains. The model can be open sourced and extended. Deployments can be reused and extended. Software and hardware can be adjusted and replaced. Governance functions can be built in and audited. This is a transition from an isolated breakthrough to a system that can actually be implemented in the real world.
This ecosystem did not emerge overnight. It builds on infrastructure “tailwinds” accumulated over the years since 2017. In recent years, China has regularly invested in data centers and computing centers, gradually forming a nationwide integrated computing layout centered on the “Data in the East, Computing in the West” strategy. The national plan established eight major computing hubs and 10 data center clusters to direct computing demand from the east to the central and western regions.
According to public information, China intends to invest in the continued expansion of energy capacity. China’s total computing power is approximately 1590 EFLOPS in 2025, ranking among the top in the world. According to Chinese sources, intelligent computing capacity tailored to AI training and deployment is expected to grow by about 43% year-on-year, far outpacing general-purpose computing. At the same time, the data center’s average power usage efficiency (PUE) decreased to approximately 1.46. This shows increased efficiency and provides a solid hardware foundation for large-scale AI. Energy is clearly a key focus.
If the 2017 “New Generation AI Development Plan” was primarily about setting direction and laying the foundations, the August 2025 “AI+” Action Plan clearly shifted the focus to large-scale deployment and tight integration. This represents a different direction to pursue than AGI. The advent of R1 provided the “lift” that was missing at the engineering and ecosystem level. This was the catalyst that systematically revitalized the computing, energy, and data infrastructure that had already been built.
As a result, in the year following the R1 release, China’s AI development accelerated along two main paths. First, AI has become more deeply embedded in industrial processes, moving beyond chatbots to agents and workflows. Second, there has been a greater emphasis on autonomous and controllable AI systems, reflected in more flexible training pathways and increasingly localized deployment strategies.
In retrospect, the real turning point was not the increase in the number of models, but the fundamental change in the way open source models were used. Open source has gone from being a choice of option to being the default assumption in system design. Models have become reusable and configurable components within large engineering systems.
Looking to the future and looking back at the past
From DeepSeek to “AI+”, China’s path to 2025 has not been about chasing peak performance. It’s about building a practical path organized around open source, engineering efficiency, and scalable delivery, and this path is already starting to take on a life of its own.
Resource constraints did not limit China’s AI development. In some ways, they have reshaped its trajectory. The release of DeepSeek R1 acted as a catalytic event, triggering a series of responses across the domestic industry and accelerating the formation of a more organically structured ecosystem. At the same time, this change has created significant opportunities to continue domestic research and development. As this ecosystem matures, its long-term impact, and how the global AI community can engage with China’s increasingly autonomous AI ecosystem, will be key questions for future discussions.
