China's Open-Source AI Leadership
China is surpassing the US in downloads of open AI models. New data from MIT and Hugging Face, cited by the Financial Times, shows that 17% of global downloads of "open" AI models originate from China, compared to 15.8% from the US. This shift has direct implications for standards, security, open-source ecosystems, and regulation.
Source: Own Representation
China's Rise in Open AI
Data from MIT and Hugging Face, published by the Financial Times, indicates that China led the US in downloads of new open AI models last year with 17%, compared to the US's 15.8%. The Financial Times cites China's aggressive release strategy by vendors as the main driver, with DeepSeek and Alibaba's Qwen mentioned as examples of high download and distribution rates. Also The Washington Post and other major media outlets confirm the growing popularity of Chinese open models on platforms such as Hugging Face.
Two model families, Qwen and DeepSeek, dominate the open landscape. Qwen 2.5 is considered widely used; the official Qwen series is freely accessible in various sizes from 0.5B to 72B, covering much of practical needs. The popular Qwen2.5-1.5B-Instruct variant is a commonly used, lightweight work model. DeepSeek releases its V and R series as open-weight models focused on efficiency; current releases and upgrades appear regularly on Hugging Face, as DeepSeek-V3 and DeepSeek-V3.1. Reuters documents these ongoing open releases and upgrades, which also put price pressure on competitors, such as the release of Alibaba's AI model intended to surpass DeepSeek-V3, or DeepSeek's update of the R1 Reasoning model.
In practice, open-source is often spoken of, although most widely distributed models are open-weight. This means the weights are accessible, but training data and full reproduction code often are not. This limits traceability and auditability, a point that analyses of the Chinese AI landscape explicitly emphasize. The same MERICS report mentions that Hugging Face has been intermittently blocked in China, which does not prevent visibility but affects infrastructure distribution. Numerical analyses from Hugging Face also show that Chinese actors are particularly strong in the segment of large open models.
Source: chat-gpt-deutschland.de
The complexity and interconnectedness of modern open-source AI models.
USA vs. China: Strategies
While the US continues to dominate private AI investment ($109.1 billion in the US vs. $9.3 billion in China in 2024), many US labs opt for more closed release models, as the AI Index Report 2025 shows. However, this is changing in nuances: even OpenAI now offers selected open-weight models, which can be seen as a reaction to the open push, as seen on the OpenAI website . In China, firms and governmental bodies are strengthening the course of broad disclosure. For example, Xiaohongshu/Rednote published its own open LLM in 2025, representing a wave of new open releases, as Reuters reports. As early as 2023, Business Standard reported on the 'hundred models race' and China's high model output, anticipating today's download statistics.
Source: t3n.de
The practical development and potential impact of open-source AI models.
Practical Implications
If open foundational building blocks primarily come from China, their security specifications, licensing models, and moderation logics shape global practice – from startup MVPs to enterprise integrations. This makes the auditability of data lineage, safety fine-tuning, and usage rights central; especially open-weight releases without data disclosure are ambivalent here, as the MERICS report highlights. At the same time, influence shifts: those who provide the most frequently used open base models de facto set standards for tokenizers, tooling, and evaluation pipelines. Analyses from the Hugging Face ecosystem demonstrate how download focal points strongly steer practical model choices. Media reports and leaderboards also show that Chinese open models are regularly visible in practical benchmarks and arenas, which further fuels adoption, as The Washington Post and the LMArena leaderboard illustrate.
For communities and early-stage teams, availability and the cost-performance ratio are crucial. Qwen and DeepSeek variants score well here: they are easily integrable into common frameworks and particularly practical in small to medium sizes, explaining their rapid downstream dissemination, for example Qwen2.5-1.5B-Instruct and DeepSeek-V3.1. The release of new open models by corporations and social platforms in China, such as von Reuters berichtet, expands the toolkit for niche and domain applications. The US counters selectively with targeted open-weight releases from research and non-profits, as well as new open initiatives from established players, intensifying the competition for de facto open standards, as seen on the OpenAI website and in the AI Index Report zu sehen ist.
China leads in downloads of new open models, thereby shifting the focus in an ecosystem that catalyzes innovation, as the Financial Times notes. For developers, startups, and regulators, this means sharpening definitions of "open", specifying safety and licensing standards, and ensuring compatibility with the most used open weights without losing sight of auditability and compliance, as MERICS and Hugging Face emphasize. The US remains financially strong but must decide how much openness it strategically allows if the open infrastructure is increasingly built elsewhere, as the AI Index Report and OpenAI show.
Source: user-added
China's ambitions in AI development are reflected in the rapid integration of technology into all areas of life.