The Disconnect Between AI Hype and Developer Usage

Hugging Face data reveals a significant gap between the AI models capturing headlines and those actually deployed by developers. The platform analyzed its 25 most-downloaded models against its 25 most-liked models, finding almost no overlap—only a single model appeared on both lists. The pattern is stark: lighter, older models dominate in actual usage. All-MiniLM-L6-v2, a model from 2021, reached 1.55 billion downloads in the first seven months of 2026 despite accumulating just over 5,000 likes. Meanwhile, not a single 2026-released model made the top 25 downloads, while 13 of the 25 came from 2022. Parameter size tells a similar story. Models under 1 billion parameters account for 83% of all historical downloads, while massive models exceeding 100 billion parameters represent just 1%. Even considering only 2026 downloads, extremely large models—those with more than 70 billion parameters—comprised merely 3% of the total. Chinese AI labs have released numerous frontier models with enormous parameter counts, including Moonshot's Kimi K3 at 2.8 trillion parameters. While these models generate considerable attention in Silicon Valley, download rates tell a different story. Kimi K3 was downloaded roughly 60 times per like received, signaling limited production adoption. Alibaba's Qwen model series offers a different approach by providing models across multiple sizes. This range appears to have positioned Qwen as a standard tool in developer workflows, resulting in approximately 2 billion downloads in 2026—roughly 55 times higher than Moonshot's figures. The trend extends to enterprise deployment strategies. Companies like Pinterest explicitly adopt a model-agnostic approach, selecting from proprietary systems, open-source alternatives, and commercial offerings based on economic and performance considerations rather than market visibility.