Open-Source AI Models Claim Growing Share of Production Workloads

The AI industry's attention may be focused on frontier models from Anthropic and OpenAI, but production workloads are increasingly running on open-source alternatives. Chinese open-weight models now represent 41% of downloads on Hugging Face this spring, surpassing U.S. models for the first time, according to TechCrunch reporting.

The shift extends beyond download metrics. On OpenRouter, the six most popular models are all open releases from Chinese companies including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Anthropic's Claude Opus 4.7 ranks seventh. Data from Vercel shows open-weight models handled nearly one-third of AI requests on the platform in June, absorbing much of the volume-heavy infrastructure work while closed models serve as a premium layer for specialized tasks.

Why it matters

This trend challenges the assumption that the most capable frontier models will dominate production AI. If enterprises increasingly deploy customized open models for cost and control reasons, the business case for multi-billion-dollar training runs becomes less clear. The data suggests a bifurcating market: frontier models for experimentation and high-value tasks, open models for scaled production workloads.

The economics of model ownership

Hugging Face CEO Clem Delangue told TechCrunch that companies are reconsidering the cost of scaling with closed frontier models. "If you're an AI company or a technology company, you don't want to outsource your core capabilities to another company, to a black box API that you don't control, don't have any visibility on, and don't really have any sort of ownership," Delangue said.

The platform now hosts nearly three million public models and one million public datasets, with a new repository created every seven seconds. Half of Fortune 500 companies use Hugging Face to deploy private or open-source models, according to Delangue. This activity points away from a "one model to rule them all" scenario toward enterprises running multiple customized models for specific use cases.

Microsoft CEO Satya Nadella echoed similar concerns about vendor lock-in, arguing that enterprises should maintain control over their data and learning infrastructure. He criticized restrictive terms on model distillation and providers reserving rights to learn from customer interaction data.

Chinese labs accelerate open releases

The rise of open models coincides with increasingly capable releases from Chinese AI labs. Beijing-based Z.ai recently released GLM-5.2, an open-weight model that competes with Anthropic's latest releases on agentic coding and security vulnerability identification. These models undercut the economics of proprietary AI by offering comparable performance at lower deployment costs with greater customization flexibility.

The safety debate intensifies

The trend has reignited debate over whether powerful models should be openly available. Anthropic CEO Dario Amodei has argued that scaling open model weights could become dangerous because they're difficult to control once released.

Delangue takes the opposite view. "The biggest risk in AI is concentration of power," he said. "The way you make the world safer, in my opinion, is by leveling up the playing fields and creating transparency on these models." He argues that transparency helps defenders patch cybersecurity risks and that keeping models closed simply concentrates technology in fewer hands while reducing visibility into how systems work.

The details were first reported by Rebecca Bellan at TechCrunch.