Chinese open-weight models captured 41% of downloads on Hugging Face this spring and now fill the top six spots on OpenRouter's live popularity leaderboard, while Anthropic's Claude Opus 4.7 sits in seventh. Hugging Face CEO Clem Delangue argues the shift is permanent, and Microsoft CEO Satya Nadella is warning enterprises not to get locked into a single model provider at all.

For several weeks this summer, the AI industry could not stop talking about Anthropic's newest models and Washington's fight over who gets to use them. Almost nobody was talking about what had quietly taken over the infrastructure underneath.

Chinese open-weight models captured 41% of downloads on Hugging Face this spring, according to the platform's own data, overtaking American models for the first time in the platform's tracked history. On OpenRouter, the marketplace developers use to route API traffic across providers, the six most popular models right now all come from Chinese firms, including Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai. Claude Opus 4.7, the model much of the industry spent June obsessing over, sits in seventh place.

"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," Hugging Face CEO Clem Delangue said on a recent episode of TechCrunch's Equity podcast.

A Third of Production Traffic Already Runs on Open Weights

The shift is not just a download counter. Data from Vercel's AI Gateway shows open models handled nearly a third of AI requests on its platform in June, positioning them as the workhorse layer for high volume applications while closed frontier models increasingly operate as the expensive option reserved for the hardest problems.

Those platforms capture only a slice of the market. Sessions hosted directly by OpenAI and Anthropic through their own apps and enterprise deals do not show up in OpenRouter or Vercel's numbers, and likely account for the bulk of what those two companies actually serve. But the trend inside the slice that is visible is unambiguous, and it raises a question every practitioner building on these systems eventually has to answer: how much does frontier model quality matter if most production traffic ends up running on cheaper, customizable alternatives instead?

Hugging Face itself is not a neutral bystander in that debate. It is the platform most open models are hosted, shared, and downloaded from, and Delangue has an obvious commercial interest in open weights winning. His numbers, though, come from Hugging Face's own usage logs rather than a survey, and they describe a platform with real scale: almost 3 million public models, over 1 million public datasets, and a new repository created every seven seconds. More than 30% of the Fortune 500 now maintain verified accounts on the platform, and Delangue says half of all Fortune 500 companies use Hugging Face to deploy either their own private models or open source ones.

The Ecosystem Flipped From Industry-Led to Community-Led

Hugging Face's own research, published in its "State of Open Source" report this spring, shows the shift goes deeper than any single leaderboard. Industry's share of overall open model development fell from roughly 70% before 2022 to about 37% in 2025. Independent developers with no corporate affiliation, the opposite end of that trend, rose from 17% to 39% of all downloads over the same stretch, and at times accounted for more than half of total usage.

China's pivot toward open weights happened fast. Following the viral release of DeepSeek's R1 model in January 2025, Chinese labs that had previously kept their best work closed reversed course. Baidu went from zero releases on Hugging Face in 2024 to more than 100 in 2025. ByteDance and Tencent each increased their release volume eight to nine times over. Alibaba's Qwen family alone now has more than 113,000 direct derivative models built on top of it, a number that climbs past 200,000 once every model merely tagged as Qwen-related is counted.

The releases keep coming. Most recently, Beijing-based Z.ai put out GLM-5.2, an open-weight model that excels at agentic coding and competes with Anthropic's latest models at identifying security vulnerabilities, all for a fraction of the price of a closed API. It followed a similar release from DeepSeek, whose V4 model matched several frontier benchmarks at a small percentage of what Anthropic charges for comparable usage. One enterprise customer told LDS it moved 100% of its production traffic from Claude to DeepSeek this year for exactly that reason, and its token bill data suggests it will not be the last to do so.

Enterprises Are Choosing Not to Rent Their Core Technology

Delangue frames the shift as enterprises waking up to what it costs to depend entirely on someone else's model.

"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," he said.

Microsoft chief executive Satya Nadella made a related argument earlier this month, warning enterprises against locking themselves into a single model provider and framing the terms most closed labs attach to their models as lopsided.

"While the great innovation that comes from model providers having fair use rights to train models on public data is needed, I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation, and to reserve the right to learn from customer usage and interaction data," Nadella said. "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."

Model developers have responded by shipping families that span far more sizes than a single flagship. A Qwen model exists for nearly every deployment tier, from under a billion parameters that runs on a phone up to hundreds of billions of parameters built for data center hardware. The emerging enterprise pattern, per Hugging Face's research, routes the bulk of requests, classification, extraction, summarization, translation, to small open models that cost a fraction of a frontier API, reserving the expensive closed model for the small share of tasks that actually need it.

Not Everyone Agrees Openness Makes AI Safer

The rise of freely downloadable, increasingly capable models has revived a fight over whether that openness is good for the world or dangerous.

Anthropic CEO Dario Amodei has argued that releasing powerful model weights publicly could become dangerous specifically because, once released, a lab loses any ability to control how the model gets used. Other critics have raised a related concern: open weights are easier for bad actors to repurpose for disinformation, cyberattacks, or biological misuse, since there is no API provider left in the loop to refuse a request.

Delangue rejects that framing entirely. "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." Keeping the most capable systems closed, he argues, does not remove the underlying risk, since attackers can already extract or steal weights from closed models and redistribute them regardless. It just concentrates who controls the technology while reducing the transparency that lets outside researchers find and patch its flaws.

"You don't really make it safe by keeping it behind closed doors for just a few players," he said. "You make it more dangerous because you create asymmetry of power and asymmetry of capabilities."

Neither side has a clean win here. Open weights genuinely do let more people audit a model's behavior, and they also genuinely remove the one lever, an API provider that can cut someone off, that closed labs use to stop obvious misuse after the fact.

The Decision Every Practitioner Now Has to Make

For data scientists and ML engineers, this is no longer a philosophical debate. It is a procurement decision most teams will face this year, if they have not already. A few things follow directly from the data:

Budget for a fallback path. Cost and ownership are pulling serious enterprise workloads toward open weights faster than most 2025 roadmaps assumed. Teams architected exclusively around one closed provider should expect pressure to add a second option.
Reframe the engineering question. Smaller open models are increasingly good enough for the bulk of production workloads: search, extraction, classification, routine coding. The real question shifts from "which model is smartest" to "which mix of models is cheapest per task at acceptable quality."
Treat any leaderboard as a snapshot, not a forecast. Hugging Face's own data shows the average open model's engagement peaks almost immediately after release and fades within about six weeks, rewarding whichever lab ships the next update fastest.

The Bottom Line

Strip away the leaderboard rankings and the underlying claim is straightforward: the most-used AI models in the world right now are not the ones dominating the headlines. A year of Chinese open-weight releases, stacked on top of enterprise cost pressure and a genuine ownership argument from Hugging Face's own CEO, has pushed a meaningful share of production AI onto models nobody has to rent.

That does not make Claude, GPT, or Gemini irrelevant. It makes them one option among many for workloads that no longer default to whichever lab shipped the most impressive demo. Delangue's bet is that frontier labs become a specialty product for the hardest problems, while everything else runs on models anyone can download. Whether OpenAI and Anthropic can defend the "everything else" business too is the question the rest of 2026 will answer.