What is Hugging Face?
Hugging Face is the largest open ML platform — a Hub of 1M+ model and dataset repos, a Serverless Inference API, hosted Spaces (Gradio/Streamlit/Docker), and dedicated Inference Endpoints on AWS/Azure/GCP.
Resources
Short guides, curated links, a glossary and FAQ for the Hugging Face Hub, Inference API, Spaces, and Datasets.
Short guides
Serverless Inference vs Endpoints
Serverless is a shared, rate-limited pool that lazy-loads models on first hit. Endpoints are dedicated instances you provision — predictable latency, per-second billing, no shared limits.
Why Spaces cold-start
Free-tier Spaces sleep after 48h of inactivity. The next request wakes the container (10–30s on CPU, 60–120s on Docker+GPU). Upgrade the hardware tier to stay always-on.
Reading a status page
A green dot means an endpoint is reachable from your browser. It does not guarantee the entire product surface — auth, storage, autoscaling, and Git LFS can degrade independently.
Why CORS limits browser checks
Some Hugging Face endpoints don't return permissive CORS headers, so browser checks only confirm DNS/TCP/TLS handshake for those — that's why we report reachability, not full JSON validation.
Curated links
Official status & docs
Glossary
- Hub
- huggingface.co — central registry for model, dataset, and Space repos.
- Spaces
- Hosted apps (Gradio, Streamlit, Docker, static) that demo a model or workflow.
- Serverless Inference API
- Shared, rate-limited inference pool that lazy-loads any Hub model on demand.
- Inference Endpoints
- Dedicated GPU/CPU instances you provision per model on AWS, Azure, or GCP.
- Router
- router.huggingface.co — OpenAI-compatible edge that fans out to multiple hosted providers.
- ZeroGPU
- Shared H100 pool available to Pro users via a @spaces.GPU decorator on a Space.
- transformers
- Flagship Python library for loading and running Transformer models.
- diffusers
- Counterpart library for diffusion models (SDXL, Flux, video).
- datasets
- Unified loader with streaming support for large corpora.
- PEFT
- Parameter-Efficient Fine-Tuning — LoRA, QLoRA, adapters.
- TRL
- Transformer Reinforcement Learning — RLHF, DPO, ORPO trainers.
- safetensors
- Safe, mmap-friendly weight format that replaced pickled .bin files.
- GGUF
- Quantized weight format for llama.cpp and Ollama-style local inference.
- Datasets Server
- Service that powers row previews and DuckDB queries in the dataset viewer.
- Cold start
- First-request latency after a container has been idle or shut down.
- Scale-to-zero
- Shut instance down after idle timeout — trades cost for cold-start latency.
- estimated_time
- Seconds hint in a Serverless 503 while a model loads.
- Gated model
- Model requiring accepted terms and an authenticated HF_TOKEN to download.
- HF_TOKEN
- User or fine-grained access token authenticating downloads and inference.
- Enterprise Hub
- Private, org-scoped Hub with SSO, audit logs, and BYO-cloud options.
FAQ
Is Hugging Face down right now?
The dashboards on this site poll the Hub and Serverless Inference edges every 15 seconds. Green dots = reachable from your browser. For component-level detail check status.huggingface.co.
Why is my Serverless Inference call returning 503?
A 503 with an estimated_time means the model is loading into the shared pool. Retry after that many seconds — warm models respond in <1s.
How is Spaces free tier different from paid?
Free-tier Spaces sleep after 48h and share a CPU pool. Paid tiers run always-on with dedicated CPU or GPU hardware.
Do you store my HF token?
No. huggingface.health is read-only and never asks for or stores tokens.
Why does this site show green but my job is failing?
Browser reachability only sees public endpoints. Auth, quotas, Git LFS, and per-endpoint capacity can degrade independently — check the official status page and your endpoint dashboard.