Qwen’s 151,000 Derivatives Move the AI Tollbooth

Alibaba introduced Qwen in 2023. By 2026, Hugging Face had counted more than 151,000 Qwen-based derivatives, yet Alibaba had not shown that reach could produce a profitable business. A model family could become infrastructure before its publisher secured the tollbooth.

When builders make an open-weight model family the default substrate for local inference, fine-tuning and agents, publishers lose the exclusive token toll. They compete instead through updates, hosting, tools, deployment, certification and enterprise workflows. Qwen gives Alibaba a route into that stack, but reusable weights do not guarantee cloud revenue or platform control.

Builders scaled Qwen before Alibaba proved the business

Hugging Face’s derivative count says more about Qwen’s position than a launch leaderboard does. Developers created more than 151,000 Qwen-based derivatives, giving the family the largest such foundation in the open-model ecosystem. Each downstream project represents a builder who accepted enough of Qwen’s behavior, packaging and capabilities to use it as a starting point. The count does not prove production use, but it sits closer to dependency than a benchmark score does.

Alibaba reported more than 3 billion global downloads of its open-weight models in six months, while Hugging Face recorded 418 million downloads for Google and 227 million for Meta in 2026. Those figures use different sources and windows, and a download records retrieval rather than commitment. Combined with the derivative base, the volume shows Qwen circulating as a building block in downstream systems.

By 2024, Alibaba said more than 90,000 businesses used its models. The company later released a Qwen2.5 model designed for edge deployment on devices including smartphones, then continued shrinking and specializing models alongside its largest releases. In 2026, Alibaba put Qwen3.8 weights—including a 27-billion-parameter model—under the Apache 2.0 license. Builders received permission to reuse and alter the weights without waiting for Alibaba to approve each application.

Builders then chose hardware, libraries, deployment patterns and evaluation practices around Qwen. Alibaba opened the door with permissive releases; thousands of developers created the installed base by carrying the family forward.

Free weights make operations the paid product

Running a dependable open-weight system still requires capacity, security, updates, observability, testing and recovery. Teams that self-host control their data, versions and availability while assuming responsibility for that work. Providers earn recurring revenue by absorbing those burdens through managed model adoption.

Alibaba Cloud sold that operational relief through OpenClaw as a service. Many OpenClaw users ran the agent platform on a Mac mini, so the software did not require a cloud gatekeeper. Alibaba Cloud, Tencent Cloud and DigitalOcean still found a product in hosting it because customers would pay to shed operational work.

Alibaba Cloud’s coding tool uses Qwen 3.5 alongside models from Zhipu, Moonshot and MiniMax. By keeping developers inside an Alibaba-managed work surface, the company can benefit from model competition without making Qwen exclusive.

An emerging model-routing layer gives enterprise buyers leverage. A buyer can select Qwen for one task, another family for the next and a different host for both. The same portability expands the market for operations while preventing the model publisher from treating adoption as captivity.

Alibaba can earn more from a transaction than a token

Alibaba has a distribution advantage that a standalone model lab lacks: its existing services already contain shopping, payments, travel and local-navigation workflows. The company linked Qwen to Taobao, Alipay, Fliggy and Amap as part of its attempt to build a one-stop AI application. Qwen can coordinate an action inside Alibaba’s network rather than stop after generating an answer.

Alibaba later opened Qwen to third-party agents from brands including KFC and Mixue. Wukong extended the same logic into enterprise work by coordinating agents for document editing and research. Merchants and corporate users bring tasks that require permissions, tools and completion states; the model supplies one component of the system.

The Qwen coupon campaign showed what workflow ownership demands. Alibaba temporarily stopped issuing coupons after customer demand overloaded the campaign. The overload made capacity planning and failure handling part of product quality as soon as the assistant touched a purchase or payment.

Rivals are hiring for the last mile

Alibaba is not alone in moving beyond model distribution. Google’s 2024 partnership with Hugging Face let developers run open models on Google Cloud compute. Google Cloud later launched Agentspace for enterprises to build and deploy agents across marketing, human resources and software development. The company made open models a source of demand for compute and managed deployment.

OpenAI launched an OpenAI Deployment Company backed by more than $4 billion and acquired consulting firm Tomoro. AWS created an AI-focused organization of forward-deployed engineers backed by $1 billion in resources. Google, OpenAI, AWS and Alibaba chose different model strategies, but all funded the organizational work between a capable model and a functioning enterprise system.

Enterprise requirements forced that convergence. Corporate deployments require data access, permissions, evaluation, integration and an accountable operator. Weights cannot negotiate with a security team or repair a broken internal process. Providers are building managed platforms and human deployment capacity because enterprise engineering remains scarce even as model access becomes abundant.

Ubiquity increases influence faster than capture

Alibaba has not shown that Qwen’s reach converts into Alibaba Cloud inference revenue, paid enterprise usage or profitable platform retention. An analysis of the company’s position concluded that Qwen had made Alibaba an AI power while it had yet to turn that global popularity into a profitable business. Downloads and derivatives establish distribution; they do not identify who hosts the resulting workloads or receives the recurring payment.

Under Apache 2.0, developers can host, alter and distribute Qwen outside Alibaba’s commercial surfaces. They can use each release without buying Alibaba Cloud, adopting Alibaba’s tools or following every later update. Alibaba gains influence through diffusion but lacks the contractual control of a closed API.

Alibaba must therefore win several separate decisions. Builders first select Qwen. Operators then choose where to run it. Enterprises choose which tools and services surround it. Finance teams finally decide whether those relationships produce durable spending. The derivative and download counts document widespread model selection, while the later conversion steps remain largely unmeasured.

Local control creates policy power and policy risk

Open weights can win precisely when a hosted provider refuses a task. Hugging Face said it used GLM-5.2 on its own infrastructure for breach forensics after safety guardrails on U.S. frontier models blocked its requests. Local hosting let Hugging Face run the workload without the Chinese model’s publisher approving it.

The same independence complicates governance. Hugging Face CEO Clement Delangue said in 2024 that he was concerned about the concentration of leading open models in China. That concern places Qwen’s international reach inside model-access geopolitics: users value freedom from hosted-provider restrictions while institutions question dependence on model families developed under another country’s policy regime.

Policymakers can amplify or interrupt the compounding process. An argument against banning Chinese open-weight models warned that the United States could cede the neutral substrate where downstream innovation accumulates. A restriction can alter which model families companies officially adopt even when permissive weights continue circulating outside the publisher’s services. Alibaba cannot fully control that circulation, and policymakers cannot treat a downloadable ecosystem like a single hosted endpoint.

Alibaba must earn control at every deployment

Qwen gives Alibaba first contact with a vast builder base, while Alibaba Cloud, T-Head, Wukong and the company’s commerce services provide places to continue the relationship. Developers can still leave Alibaba’s stack, so every layer competes on cost, reliability, integration and trust.

The 151,000-plus derivatives give Alibaba a vast field of potential customers—and every builder remains free to carry Qwen elsewhere. For Alibaba, the tollbooth now sits at deployment, where lower costs, greater reliability, tighter integration and trust earn the payment.