Chinese artificial intelligence models are gaining a stronger foothold among developers worldwide, with Alibaba’s Qwen emerging as one of the most widely used open AI model families and other Chinese systems including DeepSeek, Kimi, Zhipu AI and MiniMax rapidly expanding their global reach.

The growth is occurring while the United States continues to impose restrictions on China’s access to advanced chips and AI technologies. Rather than limiting the global adoption of Chinese models, those measures appear to be coinciding with an acceleration in the use of Chinese open-source and open-weight systems by developers looking for models that can be downloaded, customized, and deployed at relatively low cost.

Alibaba’s Qwen reportedly recorded more than 3 billion downloads worldwide over the six months through August, according to an August 14 report from Hugging Face, an open-source AI platform. That put Qwen ahead of open models from Meta and Google on the platform during the period.

Google’s open models recorded 418 million downloads, while Meta’s reached 227 million, according to the report.

The scale of Qwen’s lead is notable because downloads provide an indication of how developers are choosing the underlying technology for applications, fine-tuning, and deployment. Unlike proprietary systems that remain controlled by their developers, open-weight models can be downloaded and adapted, allowing companies and individual developers to modify them for specific applications.

Hugging Face described Qwen as “one of the largest foundations of the open AI ecosystem” and said it had become embedded in the workflow developers use when deciding which models to fine-tune and deploy.

Alibaba said the Qwen family now includes more than 460 open-source models, while more than 300,000 derivative models have been developed from them.

The breadth of that ecosystem could prove more important than the popularity of any individual model. Every derivative model represents another application, modification, or deployment built on the underlying technology, potentially expanding Qwen’s influence well beyond Alibaba’s direct users.

Tian Feng, former dean of SenseTime’s Intelligence Industry Research Institute, said the adoption figures showed that Chinese open models were gaining traction despite U.S. export controls.

The trend is also visible in the scale of models being released by Chinese AI laboratories.

According to Hugging Face’s analysis cited in the report, in almost every month of 2026 the largest and most capable open model released by a Chinese laboratory was larger by parameter count than the largest model released by a U.S. laboratory. China’s monthly maximum ranged from 754 billion to 2.78 trillion parameters, while the U.S. ceiling remained below 130 billion in five of the seven months examined.

Parameter count is not a direct measure of an AI model’s performance, but the figures indicate the scale of resources Chinese laboratories are putting into open models.

The Chinese AI industry is also continuing to release new tools aimed at developers.

Zhipu AI launched GLM-5.3 on Friday, describing the open-source release as a way to make security capabilities available to developers globally.

DeepSeek released Harness on Thursday, its first agent runtime framework. The system enables AI models to interact directly with software repositories, locate files, modify code, run tests, and repeatedly correct errors.

Moonshot AI has also fully open-sourced Kimi K3, which has 2.8 trillion total parameters and is currently the largest open model by parameter count, according to the information provided.

These releases show that competition is increasingly moving beyond chatbot applications toward the underlying infrastructure developers use to build AI products.

Chinese companies are also seeking to close the performance gap with proprietary systems developed by U.S. companies such as OpenAI and Anthropic. Models including Qwen, DeepSeek and Kimi are now being positioned as alternatives for developers who want access to advanced capabilities without being locked into a closed commercial platform.

That is creating pressure on U.S. technology companies to expand their own open-model offerings. Meta and Nvidia have released new open models in recent weeks, adding to the competition for developers and researchers who prefer models that can be downloaded and modified.

The shift toward open weights could have implications well beyond AI model developers. More accessible models can increase competition across cloud computing, semiconductors, applications and AI services by giving businesses more choices over the systems they use.

Tian said greater availability of open-weight models could stimulate innovation, reduce costs and make advanced AI capabilities more accessible and adaptable.

The debate over open AI models has also reached policymakers in Washington.

Companies and institutions including Nvidia, Microsoft, IBM, Meta, OpenAI, Cisco, Dell Technologies, GitHub and Hugging Face have signed a joint statement urging U.S. policymakers to support open-weight AI models rather than imposing restrictions that could limit their development or use.

The companies noted that U.S. leadership in AI depends on maintaining an open ecosystem across the technology industry.

Nearly 200 Silicon Valley startups operating under the Little Tech Association have also written to the U.S. government opposing restrictions on American companies’ use of Chinese open-weight AI models.

The concern among some U.S. startups is that restricting access to capable Chinese models could increase costs and limit their ability to experiment with AI technology, particularly for smaller companies that lack the resources to build models from scratch.

The growing international use of Chinese models is also extending into emerging markets.

Liu Gang, chief economist at the Chinese Institute of New Generation Artificial Intelligence Development Strategies, said Chinese open models were becoming part of the global AI supply chain and helping expand access to AI in developing economies.

Alibaba has been distributing Qwen models to enterprise customers in Southeast Asia and Africa through its cloud platform. The Qwen family supports text and multimodal applications and covers 119 languages and regional dialects, according to the Qwen team.

The potential impact is high in markets where businesses and institutions may not have the financial resources to build or license the most expensive proprietary AI systems.

Chinese open models are increasingly being used in areas including agriculture, education and healthcare, according to Liu, who said their adoption was helping narrow the digital divide across developing economies.

The latest usage data points to the scale of the shift.

Global AI model usage reached 69 trillion tokens during the week of August 3-9, according to calculations by National Business Daily based on OpenRouter data. Weekly usage increased 21.48% from the previous week.

Chinese models accounted for 34.25 trillion tokens during the period, up 21.76% week-on-week. They surpassed U.S. models for the 15th consecutive week, giving Chinese systems the largest share of global AI model usage in the period.

The figures should not be interpreted as a definitive measure of model quality or overall market share, since token usage varies according to model architecture, application, and how developers interact with different systems. But they point to a growing level of engagement with Chinese AI models among global users.

The strategic significance is becoming clearer. China’s AI industry is not competing only to produce models that match the capabilities of U.S. systems. It is increasingly trying to establish an open development ecosystem around those models, encouraging developers to download, modify, redistribute, and build on them.

That approach could give Chinese AI companies influence beyond direct commercial revenue. If developers build thousands of applications and derivative models around Qwen, DeepSeek, Kimi, or other Chinese systems, those models can become embedded in software infrastructure even when the underlying developer is not directly paying the Chinese company.

Now, the trend presents a more complicated challenge for the U.S. than simply restricting China’s access to advanced chips. Export controls can limit access to some hardware and technologies, but they do not necessarily prevent developers elsewhere from downloading and using open models that have already been released.

Alibaba is stepping up its challenge to Meta in the global open-weight artificial intelligence market, launching a smaller model designed to run on consumer devices while releasing the weights of its most powerful system for developers to download and deploy.

The move puts Alibaba directly in competition with Meta as the U.S. technology giant attempts to regain ground in open AI models and position itself as the leading American alternative to Chinese developers such as Alibaba and DeepSeek.

Alibaba launched Qwen3.8-27B last week, describing the model as capable of handling coding, professional tasks, research, and long-horizon agentic workloads. The company said the model can match the performance of another system that is 10 times larger, highlighting the industry’s push toward smaller models capable of delivering advanced performance with significantly lower computing requirements.

Alibaba also released the weights of Qwen3.8 Max, its most powerful model, allowing developers to download and run the system rather than accessing it solely through Alibaba’s cloud services.

The release of model weights has become relevant in the open-weight AI market. Weights are the numerical parameters that determine how a model processes information and generates responses. Making them available allows developers to run, modify, and fine-tune models on their own infrastructure.

However, open weights do not necessarily mean that every element of a model’s development is publicly available. The training data, training techniques, and other components used to create Qwen3.8 Max may remain undisclosed.

Alibaba has already established a strong position in the open-weight market, with Qwen becoming one of the most widely adopted model families among developers. Other Chinese companies, including DeepSeek and Moonshot AI, have also gained significant traction.

Meta was an early major participant in open AI models through its Llama family, but Chinese laboratories have rapidly expanded their presence.

The latest Alibaba release comes shortly after Meta announced plans to open-source its most powerful AI model and introduce new systems designed to operate on laptops.

Meta has also unveiled Muse Glimmer, a family of models designed for consumer computers, as it seeks to strengthen its position against Chinese open-weight models while maintaining an alternative to the more proprietary strategies pursued by OpenAI and Anthropic.

“Meta’s own re-embrace of open weights … was itself a response to two years of Chinese labs … taking a large share” of the open-weight market, Nick Patience, AI lead at the Futurum Group, told CNBC.

The competition is being measured by developer adoption rather than simply by benchmark scores.

Hugging Face, a major repository for downloadable AI models, said last week that Qwen-based models had generated 151,448 derivatives. That figure represents models built from Qwen systems after developers download and adapt them.

According to Hugging Face, Qwen’s footprint was 2.6 times that of Meta’s models.

That developer ecosystem could become one of the most valuable assets in the AI industry. A model that becomes a foundation for thousands of derivative systems can gain influence across applications, companies, and industries without its original developer having to provide every user with direct access to the underlying model.

“The company which can offer the most capable open weights models will move ahead in this race,” Neil Shah, co-founder at Counterpoint Research, told CNBC.

“Alibaba aims to become this undisputed leader, outpacing Meta and eyeing the global market … as a strong alternative to Silicon Valley frontier-grade deployable models.”

Emergence of On-device AI

Alibaba’s decision to introduce Qwen3.8-27B for consumer hardware also points to another emerging battleground: on-device AI.

Most advanced AI systems have traditionally relied on large data centers containing powerful processors. Smaller and more efficient models can instead run directly on laptops, smartphones, and other devices, reducing the need to send every request to a remote server.

That approach can provide faster responses for some applications and can offer privacy advantages because certain data can remain on the user’s device rather than being transmitted to a cloud service.

It also changes the economics of AI deployment. If capable models can operate locally, developers can build AI-powered applications without paying for every interaction with a centralized cloud model. That could accelerate adoption across consumer electronics, enterprise software and specialized devices.

Counterpoint’s Shah described on-device deployment as the “next battleground” for AI models, while Patience said Alibaba had developed an advantage across several areas, including open-weight systems and on-device AI.

“Alibaba has made Qwen the most credible non-US model family to build hardware relationships around, in China and in the open-weight developer community globally,” Patience said.

The shift toward smaller models also has strategic implications for the wider AI industry. As model capabilities improve, developers are increasingly looking beyond massive systems that require enormous data-center infrastructure and toward models that can deliver strong performance with lower computing demands.

For Alibaba, this creates an opportunity to expand Qwen’s reach beyond cloud platforms and into the hardware ecosystem. A model that becomes embedded in laptops, smartphones, and other consumer devices could establish a much broader distribution network than one that is accessed primarily through centralized data centers.

For Meta, the challenge is becoming more immediate. Its Llama models gave the company an early advantage in open AI, but Chinese competitors have demonstrated that open-weight adoption can shift rapidly when developers find models that combine strong performance, low deployment costs and broad accessibility.

The battle is therefore evolving from a contest over which laboratory produces the most powerful AI model into a race to establish the preferred foundation for developers and device manufacturers.

Alibaba’s latest releases show that it intends to compete on both fronts. By making its most powerful model’s weights available while developing smaller systems capable of running on consumer hardware, the company is seeking to extend Qwen’s influence from the developer community into the devices where AI will increasingly operate.