Alibaba's Qwen becomes world's most downloaded open AI model

Alibaba's Qwen has become the world's most-downloaded open AI model family, surpassing rivals from Meta and Google as its cumulative global downloads topped 3 billion, underscoring the growing reach of Chinese open-source AI among developers worldwide.

A report on the state of open models published by AI platform Hugging Face showed Qwen models recorded about 2.045 billion downloads on the Hugging Face Hub in 2026, compared to about 418 million for Google's models and 227 million for Meta's.

Those figures cover only downloads from Hugging Face and exclude other platforms — including China's ModelScope — meaning that they differ from Qwen's more than 3 billion cumulative global downloads across all platforms.

The numbers highlight a shift in the global open-source AI race, where Chinese models are moving beyond competing on benchmark performance and cost to building increasingly large developer ecosystems, challenging US dominance in a field long anchored by Meta's Llama family.

Alibaba has released more than 460 Qwen models as open source, which have spawned more than 300,000 derivative models, according to the data.

On Hugging Face alone, developers have created 151,448 models derived from Qwen, 2.6 times the number derived from Meta models and 4.7 times the number of repositories associated with the Llama family, the report showed. Google's models have generated 82,506 derivatives.

The rapid adoption of Qwen reflects intensifying competition over which foundation models become the underlying infrastructure for AI applications. Developers can take open models, fine-tune them for specific industries and languages, compress them to run on personal computers or other devices, and incorporate them into new applications.

That ecosystem effect is increasingly important as the AI race expands beyond the handful of companies capable of training frontier models. A widely adopted open model can become the foundation for thousands of downstream products even when the original developer does not directly operate those applications.

The Hugging Face report also pointed to a widening gap between Chinese and US laboratories in the scale and licensing of open models.

The largest open models released by leading Chinese AI laboratories in 2026 generally had more parameters than those released by their US counterparts, according to the report.

The monthly upper end for models from Chinese laboratories ranged from 754 billion to 2.78 trillion parameters, while the largest US open models remained below 130 billion parameters during most months, the report showed.

Parameter count alone is not a measure of model intelligence or efficiency, particularly as mixture-of-experts architectures can contain hundreds of billions or trillions of total parameters while activating only a fraction for each query.

Chinese developers have also generally adopted permissive licenses that allow companies and developers to modify and commercialize their models.

Of the 178 Chinese models with more than 20 billion parameters tracked in the report, 59 percent used the Apache 2.0 license and 22 percent used the MIT license. None carried restrictions limiting them to non-commercial use, according to the report.