The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+

Open source has become the dominant and popular approach for Chinese AI organizations to achieve large-scale deployment and integration. By openly sharing models, research papers, and deployment infrastructure, these organizations are leveraging intraorganizational and global community gains to integrate AI into industrial processes and workflows.

China's Organic Open-Source AI Ecosystem

An organic open-source AI ecosystem is emerging in China, characterized by a collaborative approach where established tech giants, startups, and application-focused companies contribute to a shared pool of artifacts. This collaboration extends globally, with DeepSeek and Qwen ranking as the most followed and fourth most followed organizations on Hugging Face, respectively.

Strategic Shifts of Established Players

Major Chinese technology companies have transitioned their strategies to treat open source as a core infrastructure play:

  • Alibaba: Positioned Qwen not as a single flagship model but as a family of models covering various sizes and modalities. By mid-2025, Qwen became the most derivative-heavy model on Hugging Face, with over 113,000 models using it as a base and over 200,000 repositories tagging it—surpassing Meta's Llama (27k) and DeepSeek (6k).
  • Tencent: Moved from integrating external open-source models (like DeepSeek R1) into consumer products to releasing its own capabilities. Since May 2025, Tencent has accelerated open releases under the Tencent HY (formerly Hunyuan) brand, focusing on vision, video, and 3D.
  • ByteDance: Employs an "AI application factory" approach, selectively open-sourcing high-value components. Key contributions include UI-TARS-1.5 for multimodal UI understanding, Seed-Coder for data-centric code modeling, and the SuperGPQA dataset for reasoning evaluation. Its AI application, Doubao, surpassed 100 million daily active users (DAU) in December 2025.
  • Baidu: Shifted from prioritizing closed models to re-entering the ecosystem via free access and open releases, such as the Ernie 4.5 series. Baidu is also investing in its open-source framework, PaddlePaddle, and its AI chip, Kunlunxin (which IPO'd on January 1, 2026).

Startup and Application-Layer Momentum

Startups and application-first companies have rapidly adapted to the availability of low-cost, high-reasoning open-source models:

  • AI Startups: Moonshot (Kimi K2), Z.ai (GLM-4.5), and MiniMax (MiniMax M2) have released advanced open-source models that earned places on AI-World's milestone rankings. Z.ai and MiniMax announced IPO plans in late 2025.
  • Application Companies: Firms like Xiaohongshu, Bilibili, Xiaomi, and Meituan have begun training and releasing in-house models to tune AI around specific business needs rather than relying on external providers.
  • Research Institutions: BAAI and Shanghai AI Lab have pivoted toward strengthening the ecosystem's foundations by developing toolchains, evaluation systems (OpenCompass), and data platforms (OpenDataLab, FlagOpen).

Foundations for the Future: Infrastructure and "AI+"

China's AI ecosystem is moving from isolated breakthroughs to a system capable of real-world execution. This is supported by a long-term national infrastructure strategy and a shift in policy focus.

Compute and Energy Infrastructure

The ecosystem is built on the "East Data, West Compute" strategy, which established 8 major compute hubs and 10 data center clusters. As of 2025, China's total compute capacity is approximately 1,590 EFLOPS. Intelligent compute capacity—specifically for AI training and deployment—is expected to grow by roughly 43% year-over-year. Additionally, the average data center power usage effectiveness (PUE) has decreased to approximately 1.46.

The "AI+" Action Plan

While the 2017 "New Generation AI Development Plan" focused on foundations, the August 2025 "AI+" action plan shifted the focus toward large-scale deployment and deep integration. This strategy prioritizes embedding AI into industrial processes, agents, and workflows over the pursuit of AGI. The release of DeepSeek R1 acted as a catalyst, activating the existing compute, energy, and data infrastructure to make open-source models a default assumption in system design.

Conclusion: A Practical Path to Scalable Delivery

China's AI trajectory in 2025 was defined by engineering efficiency and scalable delivery rather than chasing peak performance. Resource constraints reshaped the development path, leading to an organically structured ecosystem where models are treated as reusable and composable components within larger engineering systems.

Sources