Learn Hugging Face Hub

Hugging Face is the GitHub of machine learning. A public hub hosting 750,000+ pretrained models, with standardized weights, tokenizers, configs, and ready-to-run inference code. If a paper releases a model, it's probably on the Hub. The accompanying `transformers` Python library lets you load any of them and run inference in three lines.

The transformers library

`transformers` from Hugging Face is the de facto Python interface to most NLP/CV/audio models. The `pipeline()` helper bundles a model + tokenizer + post-processing into a callable Python function. You name the task, it picks a reasonable default model, and you can swap in any Hub model by passing `model=name`.

What lives on the Hub

• Models — pretrained checkpoints with weights + tokenizer + config
• Datasets — millions of curated training/eval datasets, streamed via `datasets` library
• Spaces — free demo apps (Gradio/Streamlit) running someone's model
• Tasks — taxonomy: text-classification, generation, ASR, vision, etc.
• Model cards — README per model with license, biases, intended use, evaluation

Self-host vs. Inference API vs. Endpoints

Three deployment options: (a) download weights and run locally — full control, your hardware. (b) Hugging Face Inference API — free tier for tinkering. (c) Inference Endpoints — paid managed hosting with auto-scaling. For prod with high throughput, you'll likely self-host on GPU or use a specialized vendor (Together, Replicate, Anyscale).