Alibaba Reports Over 3 Billion Downloads of Qwen AI Models

In the past six months, the number of downloads for Alibaba's Qwen AI models has surpassed 3 billion, according to a statement from the company reported by Bloomberg.

The Chinese tech giant has made over 460 neural networks publicly available, leading developers to create around 300,000 derivative models based on these.

Insights from Hugging Face Data

Just a day before Bloomberg's report, Hugging Face released a six-month report on the status of open models. Their calculations reveal that Qwen has been downloaded 2.05 billion times this year, with the number of derivative repositories reaching 151,448, compared to 82,506 for Google.

Source: Hugging Face.

According to Hugging Face, downloads of Google models in 2026 totaled 418 million, while Meta's reached 227 million. It's important to note that these figures only reflect activity within the Hugging Face ecosystem, excluding API requests, private deployments, and other distribution channels.

The authors of the report emphasized that these statistics should not be taken as indicators of market share or actual commercial use.

Qwen’s derivative repositories are increasing daily by 180 to 210. Of the 28,531 GGUF conversions of these models available on the platform, only 54 were created by Alibaba; the rest were developed by the community.

"Qwen has become a standard part of the workflow for developers choosing which model to fine-tune and deploy," noted Hugging Face.

Factors Behind Qwen's Success

Experts attribute Qwen's leading position to three main factors: regular updates, a range of models catering to various scales—from sub-billion versions to Qwen3.8-Max (with 2.4 trillion parameters)—and the Apache 2.0 license, which allows for modifications and commercial use without restrictions.

The diversity of the model lineup has proven to be a crucial factor. Hugging Face data indicates that models with fewer than 1 billion parameters account for 83% of all downloads in the platform’s history, while those exceeding 100 billion parameters comprise only 1%.

Laboratories focused on large language models (LLMs) are lagging behind; for instance, Moonshot AI, which rarely releases models under 70 billion parameters, managed only 37 million downloads in a year—approximately 55 times less than Qwen.

Alibaba's models are also leading in the local deployment segment, with their GGUF builds being downloaded 39.6 million times monthly, compared to Gemma's 20.8 million and Llama's 7.5 million.

The Landscape of Open AI

The report indicates a shift in the balance of power: almost every month this year, the largest open model from China has surpassed all releases from the U.S. in size. Its parameters ranged from 754 billion to 2.78 trillion, while competitors did not exceed 130 billion in five out of seven months.

In China, 59% of models with over 20 billion parameters are released under the Apache 2.0 license, 22% under MIT, with none imposing restrictions on commercial use. In contrast, U.S. developers show a different picture: 29% are under Apache/MIT, 41% under proprietary conditions, and 30% have no specified license.

Source: Hugging Face.

In the U.S., the leaders in the number of new open models are not AI laboratories but chip manufacturers: both AMD and Nvidia have each released over 200 repositories, significantly outperforming Google and Meta in this regard.

It is worth noting that in August, Techno-Nationalism author Alex Capri stated that the United States maintains its edge in the AI race not merely through individual models, but by controlling the underlying infrastructure.