This is the first blog in a series examining the historical progress of China’s open source community over the past year and its influence in shaping the entire ecosystem. Many of the advances in 2025 can be traced back to the “DeepSeek moment” in January, when Hangzhou-based AI company DeepSeek released its R-1 model.
Our first blog covers strategic changes and the explosion of new open models and open source players. The second describes architecture and hardware choices made primarily by Chinese companies in the wake of the growth of open ecosystems. You can view it here. The third analyzes the trajectory of prominent organizations and the future of the global open source ecosystem. You can view it here.
For AI researchers and developers who contribute to and rely on the open source ecosystem, and for policymakers who understand a rapidly changing environment, there has never been a better time to build and release open models and artifacts, as evidenced by the tremendous growth last year driven by DeepSeek. Notably, geopolitics is driving adoption. While models developed in China will dominate across all metrics throughout 2025, with new entrants leapfrogging each other, the Western AI community is exploring commercially deployable alternatives.
Seeds of China’s organic open source AI ecosystem
Until the introduction of R1, China’s AI industry was still centered on a closed model. Open models have existed for years, but were mostly limited to the research community or used only in niche scenarios such as privacy-sensitive applications. For most companies, these were not the default choices. Computing resources were scarce, and the question was whether to open or close.
DeepSeek’s R1 model lowers the barrier to advanced AI capabilities, provides a clear pattern to follow, and unlocks a second layer. Additionally, this release gave China’s AI development something invaluable: time. This showed that rapid progress is possible through open source and rapid iteration, even when resources are limited. This approach naturally aligned with the goals set out in China’s 2017 “AI+” strategy: to combine AI with industry as early as possible while continuing to build computing capacity in the long term.
A year after the release of R1, we see not only a growing collection of new models, but also an organic open source AI ecosystem.
DeepSeek R1: Tipping point
For the first time, a Chinese open model entered the world’s mainstream rankings and was repeatedly used as a reference point for new model releases over the next year. DeepSeek’s R1 quickly became Hugging Face’s most preferred model of all time, and most of the most preferred models are no longer developed in the United States.
However, the real meaning of R1 was not whether it was the most powerful model at the time, its importance lay in how it lowered three barriers.
The first was a technical barrier. By openly sharing inference paths and post-training methods, R1 turns advanced inference, previously locked behind closed APIs, into an engineering asset that can be downloaded, extracted, and fine-tuned. Many teams no longer need to train large models from scratch to gain powerful inference capabilities. Reasoning now behaves like a reusable module and can be applied many times across different systems. This caused the industry to rethink the relationship between model power and computing costs, and this change was especially meaningful in a compute-constrained environment like China.
The second was a barrier to adoption. R1 is released under the MIT License, making it easy to use, modify, and redistribute. Companies that relied on closed models began implementing R1 directly into production. Distillation, secondary training, and domain-specific adaptation have become routine engineering tasks rather than special projects. Once the distribution constraints were removed, the model quickly spread across cloud platforms and toolchains, and the community discussion shifted from “which model scores high” to “how to deploy, reduce cost, and integrate into real-world systems.” Over time, R1 became more than a research artifact and became a reusable engineering foundation.
The third change was psychological. When the question changes from “Can you do it?” The question, “How can we make this work?” has changed the decision-making of many companies. For China’s AI community, it was also a rare moment of sustained global attention, and a huge milestone for an ecosystem that has long been viewed primarily as a follower.
The lowering of these three barriers simultaneously means that the ecosystem begins to gain the ability to self-replicate.
From DeepSeek to AI+: Strategic realignment
As open source moved into the mainstream, natural questions arose. “How will Chinese companies’ strategies change?” Over the past year, the answer has become clear. Competition began to shift from model-to-model comparisons to system-level features.
Compared to 2024, the post-R1 release period was a time when China’s AI landscape settled into a new pattern. Big tech companies have taken the lead, startups have quickly followed suit, and companies from vertical industries are increasingly entering the space. Although their journeys have been different, a common understanding has gradually emerged, especially among leading companies, that open source is no longer a short-term strategy, but part of a long-term competitive strategy.
The number of competitive Chinese organizations releasing cutting-edge models and repositories has mushroomed. Reflecting the growth of Chinese companies’ Hugging Face repositories, the number of open releases by existing large companies has increased significantly, with Baidu going from 0 releases on Hugging Face in 2024 to over 100 in 2025, and the number of releases for companies such as ByteDance and Tencent increasing by 8-9 times. New open organizations are flooding in and releasing high-performance models, and Moonshot’s open release of Kim K2 has become “a new moment for DeepSeek.”
Releases have become strong and frequent, with high-performing models being released weekly. Newly created Chinese models are always the most liked, the most downloaded every week, and the most popular among the most downloaded new models on Hugging Face. At the top of Hugging Face’s new weekly models, you’ll see new repositories labeled by popular derivative product organization location or base model organization location.
As seen in Hugging Face’s heatmap data, there was a significant increase in open releases by Chinese companies from February to July 2025. Baidu and Moonshot have moved from a primarily closed approach to open releases. Zhipu AI’s GLM and Alibaba’s Qwen go a step further, extending from simply exposing model weights to building engineering systems and ecosystem interfaces. At this stage, just comparing raw model performance is no longer enough to win. Competition is increasingly focused on ecosystems, application scenarios, and infrastructure_._
This strategy was virtually successful. of the newly created model (
China’s AI players are not adjusting based on consensus, but rather based on constraints. What looks like collaboration is better understood as coordination under shared technological, economic, and regulatory pressures. This does not mean that the companies have entered into a cooperative partnership. Rather, the two companies began competing along similar technology foundations and engineering paths under similar computing, cost, and compliance constraints. When competition occurs between comparable system structures, ecosystems begin to exhibit the ability to spread and grow on their own. Rarely do we see the tech leaders of Zhipu AI (Z.ai), Moonshot AI, Alibaba’s Qwen, and Tencent coordinating around common questions in other countries.
Global reception and response
Positive sentiment toward open source adoption and development is growing around the world, particularly in the United States, and there is widespread recognition of how important open source leadership is to global competitiveness.
DeepSeek has been widely adopted in global markets, especially in Southeast Asia and Africa. In these markets, factors such as multilingual support, open weight availability, and cost considerations have supported enterprise use.
Western organizations often seek non-Chinese models for commercial expansion. Major releases from US organizations such as OpenAI’s gpt-oss, AI2’s Olmo, and Meta’s Llama 4 have received community engagement. Reflection AI announced an effort to build an American open-class model of the frontier. In France, Mistral has released the Mistral Large 3 family and continues to develop its open source roots.
At the same time, major releases in Western countries are built on the Chinese model. In November 2025, Deep Cogito released Cogito v2.1 as the primary open-class model in the US. This model is a finely tuned version of DeepSeek-V3. Start-ups and researchers using open weight models around the world often use models developed in China by default, even if they do not rely on them.
The American Truly Open Model (ATOM) project cites the momentum of DeepSeek and Chinese models as motivation for a collaborative effort to lead open weight model development. The project highlights the need for multiple efforts, and the study also highlights early adoption of OpenAI’s gpt-oss.
The world is responding with renewed open source enthusiasm. 2026 is poised for major releases, especially from China and the US. Of relevance are architectural trends, hardware choices, and organizational direction, which are discussed next in this series.
All data represented is provided by Hugging Face. For more information on relevant data and analysis in open source in 2025, we recommend reading Data Provenance Initiative and Hugging Face’s Economies of Open Intelligence: Tracing Power & Participation in the Model Ecosystem, aiWorld’s Open Source AI Year In Review 2025, and InterConnects’ 8 Plots Explaining the State of Open Models.
