Hugging Face

Hugging Face is an AI company and open-source community whose Hub is a central place to publish, find, and download machine learning models, datasets, and demo applications. In October 2026 the Hub lists more than 3.1 million models and 1 million datasets, from large language models (LLMs) and vision transformers to speech, robotics, and object detection checkpoints. The company also maintains widely used open-source libraries, led by Transformers.

Clément Delangue, Julien Chaumond, and Thomas Wolf founded the company in New York in 2016 as a chatbot app for teenagers, named after the hugging face emoji. Its PyTorch port of Google's BERT, published in November 2018, grew into Transformers as the company shifted from the chatbot to open-source tools. In September 2026, NVIDIA agreed to acquire Hugging Face for about $12.9 billion, in a deal expected to close in the first half of 2027, pending regulatory approval, with Hugging Face keeping its brand as an open platform.

The Hub#

Every model, dataset, and Space on the Hub is a Git repository with full version history. Large files such as weights are stored by Xet, Hugging Face's storage backend, which deduplicates data in chunks, so uploading a modified checkpoint transfers only the chunks that changed. The Hub hosts three kinds of repositories:

• Models: Model weights, configuration files, and a model card, a README whose metadata records the license, training data, evaluation results, and limitations. Many repositories store weights as safetensors, a format that loads without executing code, and gated repositories make users accept the owner's terms before downloading.
• Datasets: Data files with a dataset card and an in-browser viewer. The datasets library loads them through Apache Arrow and can stream datasets too large to download.
• Spaces: Hosted demo apps built with Gradio, Docker, or static HTML, often the quickest way to try a new model in the browser.

Inference Providers route API requests for Hub models to partner serverless providers, and Inference Endpoints deploy a model on dedicated hardware. A typical project moves through the Hub like this:

Open-Source Libraries#

• Transformers: Model definitions and pretrained weights for text, vision, audio, and multimodal tasks behind one API, including the pipeline function that runs a task in a few lines. It is now in its fifth major version.
• Diffusers: Pipelines for diffusion models that generate images, video, and audio.
• Datasets and Tokenizers: Fast data loading and processing, and fast Rust implementations of the tokenization schemes that models use.
• PEFT and TRL: Parameter-efficient fine-tuning methods such as LoRA, and post-training methods such as direct preference optimization for aligning language models.
• timm, LeRobot, and smolagents: PyTorch image models, robot learning datasets and policies, and lightweight AI agents.

Downloading Models From the Hub#

The huggingface_hub library downloads any file from a repository and caches it locally, so later calls reuse the cached copy. Ultralytics publishes YOLO26 checkpoints in the Ultralytics/YOLO26 repository, all five sizes for detection, segmentation, classification, pose estimation, and oriented boxes:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Download YOLO26n weights from the Ultralytics repository on the Hub (cached after the first call)
weights = hf_hub_download(repo_id="Ultralytics/YOLO26", filename="yolo26n.pt")

# Load the downloaded checkpoint and run detection
model = YOLO(weights)
results = model("https://ultralytics.com/images/bus.jpg")
results[0].save(filename="bus_detections.jpg")

hf_hub_download returns the path of yolo26n.pt in the local Hugging Face cache, and the saved image shows YOLO26n's detections on the bus photo.

Hugging Face vs. Related Platforms#

• GitHub: Hosts source code in Git repositories. The Hugging Face Hub uses Git too and adds what machine learning artifacts need: storage for multi-gigabyte weights, model and dataset cards, dataset viewers, and hosted inference.
• Model zoos: Framework-specific collections such as PyTorch Hub load models through one library, and the Hub serves models from many libraries, each tagged with the library that loads it.
• Model registries: Private catalogs that track a team's model versions, stages, and deployment approvals, usually inside an MLOps platform.

Real-World Applications#

• Fine-tuning open models: Teams download open-weight LLMs, adapt them with LoRA, and push the small adapter files back to the Hub instead of full copies of the model.
• Reproducible research: Papers publish their weights and datasets on the Hub, and Hugging Face's paper pages link each paper to the models, datasets, and Spaces that use it.
• Open training data: FineWeb, a Hugging Face dataset of cleaned and deduplicated English web text from Common Crawl, has grown from 15 trillion tokens at release to more than 18.5 trillion.