Chinese open-weight models accounted for 41% of downloads on Hugging Face this spring, surpassing US models, while the top six most popular models on OpenRouter are all open releases from Chinese firms including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Anthropic's Claude Opus 4.7 sits in seventh place. On Vercel's platform, open-weight models handled nearly a third of AI requests in June, absorbing the volume-heavy layer of production AI while closed models operate as the premium tier.
The numbers reframe a debate that has consumed Washington and the frontier labs for months. While OpenAI and Anthropic have pushed for tighter controls on model access and export, the developers actually shipping AI products have quietly rerouted around them. Hugging Face now hosts almost three million public models and one million public datasets, with a new repository created every seven seconds, according to CEO Clem Delangue.
Half of all Fortune 500 firms are using Hugging Face to deploy their own private models alongside open source ones, Delangue said. That is not the picture of a market converging on one model to rule them all. It is the picture of enterprises building portfolios — small, customized, often self-hosted — while treating the frontier as a specialty tool.
Key facts
- 01Chinese open-weight models accounted for 41% of downloads on Hugging Face this spring, overtaking US models.
- 02The top six most popular models on OpenRouter are all open models from Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai.
- 03Anthropic's Claude Opus 4.7 sits in seventh place on OpenRouter behind the Chinese open-weight models.
- 04Open-weight models handled nearly a third of AI requests on Vercel's platform in June 2026.
- 05Hugging Face now hosts almost three million public models and one million public datasets, with a new repository created every seven seconds.
The economic logic is straightforward. Chinese labs keep shipping open-weight models that are cheaper to deploy and easier to customize than the closed alternatives US firms have poured billions into. Beijing-based Z.ai recently released GLM-5.2, an open-weight model that excels at agentic coding and competes with Anthropic's latest releases on identifying security vulnerabilities. Every few months another release resets the price-performance floor.
“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.”— Clem Delangue, Hugging Face CEO
Delangue argues that once companies see the invoice for scaling closed frontier models, ownership starts to look better than rental. If you are an AI company or a technology company, he said, you do not want to outsource your core capabilities to a black box API you do not control, cannot see into, and do not own.
That thesis has an unlikely amplifier in Microsoft CEO Satya Nadella, who recently warned against single-provider lock-in and argued that data control should be the primary concern for enterprises using AI. Nadella said it was ironic that model providers claim fair-use rights to train on public data while imposing restrictive terms on distillation and reserving the right to learn from customer usage. If learning flows in only one direction, he said, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself.
The counterargument comes from Anthropic CEO Dario Amodei, who has said that scaling powerful open-weight models is dangerous because once released they cannot be recalled. Critics of open release argue that bad actors can more easily fine-tune open models for disinformation, cyberattacks, or biological research than they can jailbreak a closed API.
Delangue frames the tradeoff the opposite way. Transparency, he argues, lets defenders patch the cybersecurity risks that open source models can expose, while keeping powerful models closed simply concentrates the technology in a few companies and reduces visibility into how systems actually work. He also notes that API guardrails on frontier models are routinely bypassed, and that model weights themselves are a theft target.
You do not make it safe by keeping it behind closed doors for just a few players, Delangue said. You make it more dangerous because you create asymmetry of power and asymmetry of capabilities. It is a policy position with commercial alignment — Hugging Face's business is open models — but the download numbers give it a base of evidence that closed-model advocates have to answer.
There are caveats worth stating. OpenRouter, Hugging Face, and Vercel capture one slice of the market and exclude sessions hosted directly by the major labs, which likely account for the bulk of OpenAI and Anthropic's actual usage. Consumer traffic through ChatGPT and Claude apps does not show up in these leaderboards at all. Frontier revenue is not collapsing; it is being complemented, not replaced.
But the direction is set. If the highest-volume production workloads keep migrating to open, cheaper, customizable models — and if enterprise buyers keep listening to executives like Nadella on lock-in risk — then the strategic value of a frontier lead measured in months erodes fast. The labs charging premium API rates need those rates to fund the next training run, and the developer base is voting with its downloads. The race worth watching may not be who ships the smartest model in 2026, but who owns the deployment layer underneath the next million AI applications.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
