In 2026, small models dominate downloads on Hugging Face

Hugging Face observes a persistent gap between visibility and usage: the models that are popular online are not necessarily the ones that are most frequently used. In 2026, downloads are concentrated on compact architectures, often older ones. Qwen from Alibaba amassed around 2 billion downloads over the year, far ahead of Moonshot (37 million), while the timeless All-MiniLM-L6-v2 recorded 1.55 billion downloads in seven months despite only 5,156 likes.

Qwen: approximately 2 billion downloads in 2026, Moonshot 37 million

According to Hugging Face, Alibaba's Qwen series reached about 2 billion downloads in 2026, compared to 37 million for Moonshot models, resulting in a ratio of approximately 55 to 1. Hugging Face attributes this strong adoption to the variety of sizes offered by Qwen, which has allowed the series to fit seamlessly into developers' default workflows for fine-tuning and deployment.

Very large models: high parameter volumes, limited adoption

According to Hugging Face, Moonshot AI, MiniMax, Xiaomi, and Z.ai have launched few models with fewer than 70 billion parameters, some even exceeding a trillion. Moonshot's Kimi K3 model, which has 2.8 trillion parameters, recorded only about 60 downloads per like, a ratio deemed low by Hugging Face. In 2026, models with more than 70 billion parameters accounted for only 3% of downloads on the platform.

Downloads vs likes: a single intersection among the top 25

Hugging Face compared the 25 most downloaded models to the 25 models that received the most likes in 2026: only one model appears in both rankings. For the platform's researchers, a like primarily reflects the perceived importance of a release, while a download may indicate integration into an automated pipeline. No model launched in 2026 made it into the top 25 downloads, and 13 of the 25 most downloaded models date back to 2022. All-MiniLM-L6-v2, launched in 2021 by Sentence Transformers, exemplifies this phenomenon with 1.55 billion downloads in the first seven months of 2026 for only 5,156 likes. Structurally, among the repositories that publish their size, models with fewer than 1 billion parameters account for 83% of historical downloads, compared to 1% for those with more than 100 billion.

Deployments: a pragmatic choice of models

Pinterest specifies that it does not favor any particular type of model for its AI deployments: the company uses its internal models for personalization, open-source models if they offer a better cost-performance ratio, and proprietary models when they deliver better performance.