The Disconnect Between AI Model Popularity and Developer Usage
In the rapidly evolving field of artificial intelligence, the models that attract the most attention are not necessarily the ones being widely adopted by developers. This intriguing insight comes from recent data shared by Hugging Face, a platform renowned for allowing developers to publish, share, and download AI models.
Hugging Face’s analysis compared the 25 most downloaded AI models of the year to the 25 models that received the highest number of likes. Surprisingly, only one model appeared on both lists, highlighting a significant gap between the models that developers express interest in and those they actually use.
The AI community often focuses on cutting-edge frontier models, emphasizing the latest releases, benchmark achievements, and social media buzz. Developers, however, frequently opt for smaller, more stable models that are cost-effective and seamlessly integrated into existing production systems.
Hugging Face’s researchers noted, “A like says a release matters,” whereas a download indicates that a model might be “wired into a pipeline that runs on a schedule.” This distinction underscores the practical needs of developers as opposed to the hype surrounding new releases.
A prime example of this trend is the All-MiniLM-L6-v2 model, released in 2021 by Sentence Transformers. This fast, lightweight model was downloaded 1.55 billion times in just the first seven months of 2026, despite garnering only 5,156 likes. Moreover, none of the models released in 2026 made it to Hugging Face’s top 25 by downloads, while 13 of the top 25 were from 2022.
Smaller models, defined as those with fewer than 1 billion parameters, dominate the landscape, making up 83% of all-time downloads among repositories that disclose parameter counts on Hugging Face. Conversely, large models with over 100 billion parameters, like Moonshot AI’s Kimi K3, represent a mere 1% of downloads.
Even when examining downloads exclusively from 2026, models with more than 70 billion parameters accounted for only 3% of the total, a surprising statistic given the focus on massive frontier models, particularly from Chinese laboratories.
Chinese companies such as Moonshot AI, MiniMax, Xiaomi, and Z.ai have produced several high-profile models, often exceeding 1 trillion parameters. However, Hugging Face data indicates that these models see limited use compared to their smaller counterparts. For instance, Moonshot’s Kimi K3 model, with its 2.8 trillion parameters, generated significant buzz but was downloaded only about 60 times for each like it received.
In contrast, Alibaba’s Qwen model series has adopted a different strategy by offering a range of model sizes. This versatility has made Qwen a staple in developers’ “default workflow” for fine-tuning and deployment, resulting in about 2 billion downloads in 2026, significantly outpacing Moonshot’s 37 million.
This divergence between model hype and actual usage is also evident in corporate AI deployment strategies. For example, Pinterest has previously stated that it employs a “model-agnostic” approach, utilizing its own models for personalization, open-source AI models for cost efficiency, and closed models for optimal performance.
