The AI race is usually narrated through a familiar cast: OpenAI, Google, Meta, Anthropic, all chasing the smartest system. But there is a second race, measured not by benchmark scores but by how many developers actually reach for a model and build on it, and by that measure the leader is not American at all. Alibaba says its Qwen family of open-weight models has been downloaded more than 3 billion times in the past six months. According to Hugging Face’s State of Open Models report, published on 14 August, Google’s open models recorded around 418 million downloads this year and Meta roughly 227 million. Qwen is not ahead by a nose. It is ahead by an order of magnitude.
Raw download counts can flatter, so the more telling number is what developers do next. Alibaba has released over 460 Qwen models, and the community has spun those into more than 300,000 derivatives: fine-tuned, trimmed, and re-pointed at specific jobs. That is the whole logic of open weights. Unlike a closed service you rent through an API, a downloadable model can be customised, trained on private data, and embedded as the foundation of something new. Every download is a potential starting point, and Hugging Face now calls Qwen “one of the largest foundations of the open AI ecosystem,” a default choice when developers decide what to fine-tune and deploy.
Adoption at that scale tends to feed itself. More developers building on Qwen produces more specialised variants, which draw in still more developers, a snowball that is hard to reverse once it starts rolling. Alibaba has a second lever few open-source labs enjoy: a large cloud business. It is pushing Qwen to enterprise customers beyond China, across Southeast Asia and Africa, putting the models in front of companies that might otherwise default to an American option. Distribution, in this contest, is not a side effect of quality. It is the prize.
Qwen is also not alone. DeepSeek, Moonshot’s Kimi and MiniMax are all pushing capable open models out of China, trying to close the gap with the closed frontier systems from OpenAI and Anthropic. What is striking is how little the American effort to slow them down seems to bite. Export controls on chips, and even a brief block this summer on overseas access to Anthropic’s Fable 5, have not obviously dented the pace. Open weights travel through developers and repositories, not through customs, which makes them awkward to contain.
None of this means Qwen is the best model in the world, and downloads are not revenue. A company can lead the open ecosystem while others still set the capability frontier and collect the enterprise contracts. But there is a real prize in being the substrate everyone else builds on, the way an operating system quietly shapes an entire market above it. US firms have noticed: Meta and Nvidia have both shipped new open models in recent weeks. The next phase of the AI contest may hinge less on who trains the cleverest system, and more on whose model the rest of the world quietly decides to build with.