Hugging Face explained: the open-model hub, its prices and who needs it

Hugging Face is where the open model you read about in the news lives, with the dataset it trained on and the demo somebody built from it. Clément Delangue, Julien Chaumond and Thomas Wolf founded the company in 2016 with a chatbot for teenagers, and nothing of that product survives; the repository the whole field links to grew out of it.

“The open-model repository the whole field links to”

— Intokened, Our card on the service

What sits on the platform

Four things sit on the platform. Open models and datasets anyone can pull. Hosting for your own model and a demo page beside it. GPU rental for training runs. An API over all of it, which is how most production code touches the hub. Our service card keeps the feature list and the checked prices together.

What people actually download

The traffic shows what people use it for. The most downloaded model on the hub today is sentence-transformers/all-MiniLM-L6-v2, a small embedding model from 2021, with 218,138,134 downloads in the past 30 days. The most liked model, Qwen3.8-27B, holds 17,660 likes and 7.1 million downloads. The small embedding model gets pulled 30 times more often than the most admired LLM.

Downloads per like tell it from the other side: 34,412 for the embedding model, 41 for FLUX.1-dev, which collected 15,562 likes against 643,822 downloads. The likes gather on image generators, and the installs go to search infrastructure.

What it costs

Pricing runs to four tiers. Public access stays free. PRO costs $9 a month and includes $2 of inference credits, so 22% of the fee returns as compute and the rest buys private space and a visible badge. Team runs $20 per seat for private repositories, Enterprise $50 per seat with access control and audit logs. Ten people on Team pay $200 a month against $9 for one person on PRO.

Spending here gets measured in compute. Rent a GPU for a training run and the bill follows GPU hours, with the tier deciding what you may keep private rather than what you may spend.

Who needs it: a developer pulling a model into an application, a researcher publishing weights and wanting them found, a team that needs private repositories with an audit trail. Someone who wants to talk to a chatbot needs none of it, and our platform selection says which tool answers that instead.

Informational material, not a purchase recommendation. Download counts were read from the public hub API on 14 October 2026 and cover the past 30 days; plan prices were checked against the official page on 12 October.