Alibaba's Qwen AI models have surpassed 3 billion downloads, driven by extensive community development and a wide range of available models.
Alibaba has announced that its Qwen AI models have accumulated over 3 billion downloads in the last six months, marking a significant milestone for the Chinese tech giant. The company has made available more than 460 neural networks, with developers creating approximately 300,000 derivative models based on these.
Hugging Face Insights
In a recent report, Hugging Face highlighted that Qwen models have been downloaded 2.05 billion times this year alone, with derivative repositories reaching 151,448. In comparison, Google models garnered 82,506 derivative repositories. The report emphasizes that these figures are strictly from the Hugging Face ecosystem and do not account for API requests or private deployments.
Daily, Qwen sees an increase of 180 to 210 new derivative repositories. Of the 28,531 GGUF conversions available on the platform, only 54 originated from Alibaba, indicating a robust community development effort.
According to Hugging Face, "Qwen has become a standard part of the workflow for developers choosing which model to fine-tune and deploy."
Factors Contributing to Qwen's Popularity
Experts attribute Qwen's success to its regular updates, a diverse range of models that cater to different scales, and the permissive Apache 2.0 license that allows for unrestricted modifications and commercial use. Notably, models with fewer than 1 billion parameters have accounted for 83% of all downloads on the platform, contrasting with those over 100 billion parameters, which make up only 1%.
In the local deployment sector, Alibaba's models are leading, with GGUF builds downloaded 39.6 million times each month, significantly outpacing competitors like Gemma and Llama.
Shifting Dynamics in Open AI Models
The report indicates a notable shift in the landscape of open AI models, where the largest releases from China have consistently outperformed U.S. models in size this year. Chinese models have parameters ranging from 754 billion to 2.78 trillion, while U.S. models have not exceeded 130 billion in five of the past seven months.
In China, 59% of models with over 20 billion parameters are offered under the Apache 2.0 license, while in the U.S., the distribution is mixed, with 41% under proprietary conditions and 30% lacking a specified license. Interestingly, chip manufacturers like AMD and Nvidia have outpaced AI labs in the U.S. by releasing over 200 repositories each, surpassing Google and Meta in this domain.
In August, Alex Capri, author of Techno-Nationalism, remarked that the U.S. maintains its competitive edge in AI not just through individual models but by controlling the underlying infrastructure.
