The AI models getting the most attention aren’t the ones developers are actually using, according to new data from Hugging Face.
Hugging Face is a platform where developers publish, share, and download AI models. Users can also “like” models, a signal of interest, while downloads offer a picture of actual usage.
The company compared the 25 models with the most downloads this year with the 25 that received the most likes. It shows a sharp divide between the models developers are excited about and the ones they actually rely on. Only one model landed on both lists.
The AI industry tends to obsess over the newest frontier models: the biggest releases, the strongest benchmark scores, and the launches that dominate social media. Developers, meanwhile, often stick with smaller and older models that are cheap, stable, and already embedded in production systems.
“A like says a release matters,” Hugging Face’s researchers wrote. A download, by contrast, can mean a model is “wired into a pipeline that runs on a schedule.”
All-MiniLM-L6-v2, a fast, light-weight model released in 2021 by Sentence Transformers, was downloaded 1.55 billion times in the first 7 months of 2026 despite receiving just 5,156 likes. Also, not a single model released in 2026 made Hugging Face’s top 25 by downloads, while 13 of the 25 were released in 2022.
Smaller models, those under 1 billion parameters — the numerical values a model learns during training that help determine how it processes and generates outputs — account for 83% of all-time downloads among Hugging Face repositories that disclose their parameter counts. Larger models, those with more than 100 billion parameters — such as Moonshot AI’s Kimi K3 — account for just 1%
Even when Hugging Face looked only at downloads accumulated in 2026, models with more than 70 billion parameters represented just 3% of the total.
Those stats are surprising given the amount of attention that’s gone to enormous frontier models, especially from Chinese labs.
Moonshot AI, MiniMax, Xiaomi, and Z.ai are all buzzy Chinese models that have released few small models below 70 billion parameters, according to Hugging Face. Some have pushed well past 1 trillion parameters.
Moonshot’s Kimi K3, for instance, arrived with 2.8 trillion parameters and generated a ton of buzz in Silicon Valley for its coding performance and lower price. But it was only downloaded about 60 times per like it received, which is low, Hugging Face said.
Alibaba’s Qwen model series has taken a different approach, spanning a wider range of sizes. Hugging Face says that breadth has helped make Qwen part of developers’ “default workflow” for fine-tuning and deployment — and it has translated into much higher usage. Qwen models logged about 2 billion downloads in 2026, roughly 55 times Moonshot’s 37 million.
That gap between hype and usage is also showing up in how companies deploy AI. Pinterest, for example, previously told Business Insider that it takes a “model-agnostic” approach, using its own models for personalization, open-source AI models when they offer better economics, and closed models when they perform best.
