Chinese open AI models now lead downloads. Should your business use them?

Most business owners think of AI as three or four brand names you pay by the month. Underneath that, a second market has grown up fast: free, downloadable "open-weight" models that anyone can run on their own servers. This year the biggest players in that market stopped being American. That shift is worth understanding before you sign your next AI contract.

What happened

Hugging Face, the largest public library of AI models, published its State of Open Source report for spring 2026. It found that Chinese models passed U.S. models in both monthly and total downloads, reaching a plurality of 41% over the past year. Alibaba's Qwen family alone has more derivative models (versions other people built on top of it) than Google and Meta combined. The report also notes that companies such as Baidu and MiniMax, which used to keep their models closed, moved toward open releases.

A follow-up, Hugging Face's summer 2026 observations covering January to August, adds detail. The largest Chinese open releases ran from hundreds of billions to trillions of parameters, while most U.S. open releases stayed far smaller. Licensing differs too: of the Chinese releases above 20 billion parameters, most used permissive licenses like Apache 2.0 or MIT, while many U.S. releases used custom terms or declared no license at all. The same post flags that a couple of large Chinese models recently added non-commercial or revenue-share terms, so "open" does not always mean "free for any use."

TechCrunch framed the bigger question in July in a piece arguing that the real AI race may no longer be at the frontier. It reported that the six most popular models on the routing service OpenRouter at the time were open models from Chinese firms, and that on Vercel's platform open models handled nearly a third of AI requests in June. Hugging Face CEO Clem Delangue told the outlet companies want to own their core AI rather than rent "a black box API." The article also gave the other side: Anthropic CEO Dario Amodei has argued that powerful open weights could become dangerous because they are hard to control once released.

Why it matters for your business

Think of closed models (the big paid assistants) as renting a finished office, and open models as buying a building you can remodel. Renting is easier and usually better on day one. Owning gives you control over cost, data and how long the thing stays the same.

For a small or mid-size business, three points stand out:

• Cost at volume. Routine, high-volume jobs like sorting emails, tagging documents or summarizing calls do not need the most powerful model on earth. TechCrunch's reporting describes open models carrying that bulk work while premium closed models handle the harder cases. That split can cut a monthly bill significantly.
• Control of your data. An open model can run on hardware you or your provider controls, so client files never pass through a third party's servers. For law firms and anyone handling sensitive records, that is often the deciding factor.
• Origin and license risk. A model's country of origin matters to some clients, insurers and government contracts. Licenses also vary and can change between versions. Download counts measure popularity, not quality or safety, and Hugging Face itself cautions against reading its numbers as market share.

What we'd do

• Sort your AI tasks into two piles: judgment-heavy work that justifies a premium model, and repetitive work where a smaller open model is likely good enough.
• Test one open model on one repetitive task with your own real examples before committing. Measure accuracy and cost side by side with your current tool.
• Read the license for the exact version you use, not the family name. Check for non-commercial clauses and revenue-share terms.
• Ask your vendor where your data goes and which model sits under their product. Many AI tools quietly run on open models already.
• Check customer and contract rules on foreign-developed software before you deploy anything that touches regulated data.

Our take

We think the smart default for most businesses is a mix, not a single vendor. Closed frontier models are still the best choice for hard reasoning and they come with support and safety work you would otherwise do yourself. But paying premium rates for routine work, and locking every process into one provider, is a habit that will look expensive within a year. Where a model comes from matters less than whether you can test it, control it and switch away from it. Build so that you can.