Open-weight models surge past frontier AI, reshaping where real work happens

For much of the summer, the AI industry's attention was fixed on Anthropic's latest frontier models and regulatory battles in Washington over who should be granted access to advanced AI. But while industry watchers focused on the frontier, a different race was unfolding beneath the surface: developers and enterprises were increasingly adopting open-weight models from Chinese and other non-frontier labs.

The data tells a striking story. On Hugging Face, a platform and developer community for hosting and deploying open models, Chinese open-weight models accounted for 41% of downloads this spring, surpassing U.S. models. More dramatically, on OpenRouter—a service aggregating popular AI models—the top six most popular models are all open models from Chinese firms: Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Anthropic's Claude Opus 4.7 ranks seventh. On Vercel, which tracks AI infrastructure usage, open-weight models handled nearly a third of AI requests in June, while closed models operate as a higher-cost, premium layer.

Clem Delangue, CEO of Hugging Face, argues that this pattern reflects a structural change in how enterprises view AI. Rather than renting capabilities from a single provider via an API, companies increasingly want to own and control their own models. "Maybe in a few years, the frontier models will be for experimenting and [for] some really high-value tasks, and most of the production workloads will actually be powered either by private models within companies or by open source models," Delangue said. The shift has picked up steam as companies face mounting bills for scaling closed frontier models and realize they do not want to "outsource [their] core capabilities to another company, to a black box API that you don't control." Half of all Fortune 500 firms are now using Hugging Face to deploy their own private models and open-source models.

The scale of this activity is extraordinary. A new repository is created every seven seconds on Hugging Face, which hosts almost three million public models and one million public datasets. This points, Delangue argues, not to "one model to rule them all," but rather to companies using many different models, many customized for their specific use case.

Chinese AI labs have accelerated this trend by releasing increasingly capable open-weight models on a predictable schedule. Every few months, another Chinese AI company releases a powerful open-weight model that is cheaper to deploy and easier to customize than closed competitors. Most recently, Beijing-based Z.ai released GLM-5.2, an open-weight model that excels at agentic coding and competes with Anthropic's latest models on identifying security vulnerabilities.

Microsoft CEO Satya Nadella has voiced similar concerns about single-provider lock-in. "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself," Nadella said. "Therefore, it's imperative that we distribute the learning infrastructure to every firm so that they can control their own learning loop."

The rise of open models has intensified debate over safety. Anthropic CEO Dario Amodei has warned that scaling powerful open model weights could become dangerous once released, and critics argue that open models are easier for bad actors to access and misuse for disinformation or cyber or biological warfare. Delangue counters that the biggest risk in AI is concentration of power, and that "the way you make the world safer, in my opinion, is by leveling up the playing fields and creating transparency on these models." Open models allow defenders to patch cybersecurity risks more easily, he argues, and keeping models closed does not eliminate risk—it simply creates asymmetry of power and visibility.

The momentum in open-weight models reflects a fundamental shift in how enterprises are choosing to deploy AI. Rather than waiting for access to the latest frontier models from Anthropic, OpenAI, and other leading labs, developers and companies have been building and deploying alternatives—a trend that platforms like Hugging Face, OpenRouter, and Vercel now make visible at scale. Chinese AI companies have capitalized on this demand by releasing increasingly capable open models faster and at lower cost than U.S. competitors, capturing the majority of download volume in recent months.

The economic logic driving this shift is clear: closed frontier models are expensive to scale, and their licensing terms restrict how companies can use and learn from their own data. Enterprises prefer models they can customize, control, and host privately. Hugging Face's data showing that a new repository is created every seven seconds on its platform, and that half of Fortune 500 firms deploy their own models there, underscores how much technical and business activity is now happening outside the walled gardens of proprietary AI providers.

This dynamic raises a direct challenge to the "winner-take-all" narrative that has long dominated AI discourse. If most production workloads are powered by open models while frontier models serve only specialized, high-value tasks, the strategic importance and commercial advantage of being first at the frontier diminishes substantially. The real competition may now be in the middle market—where speed to deployment, cost, customizability, and domain-specific capability matter more than raw benchmark performance.

No comments yet. Be the first to share your thoughts!