The hottest AI models aren’t the ones developers actually use

The AI models getting the astir attraction aren't the ones developers are really using, according to caller information from Hugging Face.

Hugging Face is simply a level wherever developers publish, share, and download AI models. Users tin besides "like" models, a awesome of interest, while downloads connection a image of existent usage.

The institution compared the 25 models pinch the astir downloads this twelvemonth pinch the 25 that received the astir likes. It shows a crisp disagreement betwixt the models developers are excited astir and the ones they really trust on. Only 1 exemplary landed connected some lists.

The AI manufacture tends to obsess complete the newest frontier models: the biggest releases, the strongest benchmark scores, and the launches that dominate societal media. Developers, meanwhile, often instrumentality pinch smaller and older models that are cheap, stable, and already embedded successful accumulation systems.

"A for illustration says a merchandise matters," Hugging Face's researchers wrote. A download, by contrast, tin mean a exemplary is "wired into a pipeline that runs connected a schedule."

All-MiniLM-L6-v2, a fast, light-weight exemplary released successful 2021 by Sentence Transformers, was downloaded 1.55 cardinal times successful the first 7 months of 2026 contempt receiving conscionable 5,156 likes. Also, not a azygous exemplary released successful 2026 made Hugging Face's apical 25 by downloads, while 13 of the 25 were released successful 2022.

Smaller models, those nether 1 cardinal parameters — the numerical values a exemplary learns during training that thief find really it processes and generates outputs — relationship for 83% of all-time downloads among Hugging Face repositories that disclose their parameter counts. Larger models, those pinch much than 100 cardinal parameters — specified arsenic Moonshot AI's Kimi K3 — relationship for conscionable 1%

Even erstwhile Hugging Face looked only astatine downloads accumulated successful 2026, models pinch much than 70 cardinal parameters represented conscionable 3% of the total.

Those stats are astonishing fixed the magnitude of attraction that's gone to tremendous frontier models, particularly from Chinese labs.

Moonshot AI, MiniMax, Xiaomi, and Z.ai are each buzzy Chinese models that person released fewer mini models beneath 70 cardinal parameters, according to Hugging Face. Some person pushed good past 1 trillion parameters.

Moonshot's Kimi K3, for instance, arrived pinch 2.8 trillion parameters and generated a ton of buzz successful Silicon Valley for its coding capacity and little price. But it was only downloaded astir 60 times per for illustration it received, which is low, Hugging Face said.

Alibaba's Qwen exemplary bid has taken a different approach, spanning a wider scope of sizes. Hugging Face says that breadth has helped make Qwen portion of developers' "default workflow" for fine-tuning and deployment — and it has translated into overmuch higher usage. Qwen models logged astir 2 cardinal downloads successful 2026, astir 55 times Moonshot's 37 million.

That spread betwixt hype and usage is besides showing up successful really companies deploy AI. Pinterest, for example, antecedently told Business Insider that it takes a "model-agnostic" approach, utilizing its ain models for personalization, open-source AI models erstwhile they connection amended economics, and closed models erstwhile they execute best.