A wide gap has emerged between the latest and very large AI models drawing the most attention in the industry and the models widely used in real development settings. Business Insider reported on Aug. 17 that Hugging Face compared its top 25 models by downloads this year with its top 25 models by likes, and found only 1 model appeared on both lists.

On Hugging Face, likes show developer interest, while downloads indicate how much a model is used in actual services or development pipelines. Industry attention is focused on the size and benchmark performance of the latest frontier models, but developers tend to keep using smaller, older models that are cheaper and have proven stability.

A representative example is All-MiniLM-L6-v2, released in 2021. The lightweight and fast model was downloaded 1.55 billion times in the first 7 months of this year, but had only 5,156 likes. None of the models released this year made the top 25 by downloads, and 13 of the top 25 were models released in 2022.

Model size also tended to move inversely to actual usage. Based on Hugging Face models that disclose parameter counts, small models with fewer than 1 billion parameters accounted for 83 percent of cumulative downloads. By contrast, very large models with more than 100 billion parameters made up only 1 percent. Looking only at downloads this year, models with 70 billion parameters or more were about 3 percent of the total.

The gap was also pronounced among Chinese AI companies that joined the race for very large models. Moonshot AI's Kimi K3 drew strong attention for its 2.8 trillion parameters, coding performance and low price, but it logged only about 60 downloads per like.

By contrast, Alibaba pushed deep into the development ecosystem by offering its Qwen models across a wide range of sizes, from small to large. Qwen was downloaded about 2 billion times this year, about 55 times more than Moonshot AI's 37 million. Hugging Face said the strategy of providing models in various sizes made Qwen a default workflow for developers in fine-tuning and actual deployment.

The implication is that the latest performance and model size drive buzz in the AI market, but cost, stability and ease of deployment determine model selection in the field. A "model-agnostic" strategy that combines in-house models with open and closed models depending on cost and performance, as Pinterest does, also reflects the trend.