China's open-source AI ecosystem just outgrew DeepSeek: 113,000 Qwen derivatives now eclipse Meta

The global open-source AI map is being redrawn, and the center of gravity has decisively shifted. In the months since the “DeepSeek Moment” jolted the industry in early 2025, a full-blown organic ecosystem has taken root in China, and it’s no longer reliant on a single lab’s breakthroughs. The numbers are stark. Alibaba’s Qwen model family, treated less as a flagship and more as a general-purpose foundation, has now been used as the base for over 113,000 derivative models on Hugging Face. That’s more than four times Meta’s Llama and nearly twenty times DeepSeek’s own base. The collaborative flywheel is spinning, with Chinese organizations also dominating the most-read research papers on the platform.

The strategy is diverging sharply by company. Alibaba is embedding Qwen into a single stack spanning cloud infrastructure and custom silicon, making the model impossible to decouple from the compute it runs on. Tencent took the opposite route initially—quietly integrating DeepSeek into its consumer products before launching its rebranded Hunyuan (now HY) models for vision and video tasks. ByteDance, whose Doubao app quietly crossed 100 million daily active users, is taking an “application factory” approach, selectively open-sourcing high-value components like UI-TARS and Seed-Coder while keeping its product moat firmly closed.

Perhaps the most telling signal comes from an unlikely convert. Baidu, whose CEO Robin Li publicly dismissed open source for years, has capitulated. The company not only released its Ernie 4.5 series but is now connecting those models to its PaddlePaddle framework and Kunlunxin AI chips, which just filed for an IPO. This isn’t a change of heart—it’s a strategic recalibration under shared constraints of compute and cost. By opening up, Baidu can attract developers and influence standards while maintaining control at the infrastructure layer.

For policymakers and developers watching this space, the implication is clear: the open-source model isn’t just a research philosophy anymore. It’s a deployment strategy. When models, papers, and even infrastructure maps are shared openly, the goal isn’t charity—it’s large-scale integration. The question that remains isn’t who will release the next reasoning model, but which organization’s stack will become the default plumbing for the next generation of applications.