Models, datasets, and inference for open machine learning.
CodingResearch
Models, datasets, and inference for open machine learning. Browse the Hub to find models, try them in browser widgets powered by Inference Providers, and integrate the same models via SDK or API. Hub seats (Free, PRO, Team, Enterprise) are separate from pay-as-you-go inference credits.
Filter models by task and Inference Providers, open a model page, read the Model Card, then test in the widget on the right before writing code.
What to provide
Details that improve the result
Example prompt
Find a conversational text-generation model for summarization. Filter by Inference Providers, open a top model, read its Model Card for language support, then enter: “Summarize in three bullets: [paste text]” in the widget and click Generate.
If the first output is not good
Common mistakes
Use InferenceClient or the OpenAI-compatible endpoint at router.huggingface.co/v1 with Bearer auth. Pass model id and messages; append :fastest to let the router pick a provider.
What to provide
Details that improve the result
Example prompt
explorePython with InferenceClient:
client.chat.completions.create(
model="deepseek-ai/DeepSeek-V3-0324",
messages=[{"role": "user", "content": "Explain LoRA fine-tuning in five sentences."}],
max_tokens=300
)If the first output is not good
Common mistakes
Search Datasets on the Hub, read the Dataset Card, preview rows in Data Studio, then load with the datasets library or download files directly.
What to provide
Details that improve the result
Example prompt
Find a small English sentiment dataset with an open license. Open Data Studio to preview columns, then in Python: from datasets import load_dataset; ds = load_dataset("org/name", split="train[:100]")If the first output is not good
Common mistakes
Search Spaces, open a demo, interact in the browser. For your own Space, create a repo and deploy with Gradio or Streamlit SDKs.
What to provide
Details that improve the result
Example prompt
Search Spaces for a text-to-image Gradio demo. Enter “A watercolor forest at dawn, no text” and run. Note latency and output quality before integrating the underlying model via API.
If the first output is not good
Common mistakes
Create a model repo on the Hub, upload files via git or huggingface_hub, add a README Model Card, set pipeline_tag, and enable an inference widget if a provider hosts it.
What to provide
Details that improve the result
Example prompt
Create repo username/my-classifier. Upload config and weights, write a Model Card with task, training data summary, evaluation metrics, and license. Tag pipeline_tag: text-classification.
If the first output is not good
Common mistakes
Install huggingface_hub or @huggingface/inference. Use InferenceClient with provider="auto" for routing, or call task-specific methods. Copy snippets from the model widget when unsure.
What to provide
Details that improve the result
Example prompt
from huggingface_hub import InferenceClient
client = InferenceClient(api_key=HF_TOKEN, provider="auto")
image = client.text_to_image("A serene mountain landscape at sunset", model="black-forest-labs/FLUX.1-schnell")If the first output is not good
Common mistakes
On Hugging Face, prompting happens in model widgets, the Inference Playground, or your API calls. Match the prompt style to the task and model: chat messages for LLMs, descriptive text for image models, and structured inputs for task-specific pipelines.
Task: chat summarization.
Model: deepseek-ai/DeepSeek-V3-0324
Messages: [{"role": "system", "content": "Return three bullets only."}, {"role": "user", "content": "Summarize: [text]"}]
Constraints: max_tokens=200Model Cards list intended tasks, languages, limitations, and bias notes. Prompt within those bounds for reliable results.
Example
Before prompting a translation model, confirm its supported language pairs in the Model Card, then: “Translate to French: [English text]”.
Chat completion models expect role-structured messages. A system message sets format; user messages carry the task.
Example
messages=[{"role": "system", "content": "Answer in JSON: {summary, open_questions}"}, {"role": "user", "content": "Analyze this paragraph: …"}]The model-page widget uses the same Inference Providers endpoint as production. Validate prompts interactively, then copy View Code Snippets.
Example
Enter your prompt in the widget, click Generate, adjust until output fits, then copy the generated Python snippet.
Append :fastest for speed, :cheapest for cost, or :groq/:fal/etc. for a specific provider. Omit suffix to use default fastest routing.
Example
model="moonshotai/Kimi-K2-Instruct-0905:groq" for a fixed provider; model="deepseek-ai/DeepSeek-R1-0528:fastest" for auto speed.
Text-to-image models want visual descriptions; embedding models want raw text; ASR models want audio input. Check pipeline_tag on the model page.
Example
For text-to-image: “A serene mountain landscape at sunset, wide angle, no text, no watermark.”
Start on the model page
Open a model repo, read the Model Card, and use the inference widget on the right. Widgets appear when at least one Inference Provider hosts the model.
Use the Inference Playground for chat models
Compare chat completion models with the same prompt, adjust temperature and max_tokens, and pick a model before wiring it into your app.
Authenticate API calls properly
Generate a fine-grained Hugging Face token with Inference Providers permission. Pass it as Bearer hf_**** in headers or via HF_TOKEN in SDKs.
Choose the right API surface
Chat tasks: OpenAI-compatible router.huggingface.co/v1 or InferenceClient.chat.completions. Other tasks (image, speech, embeddings): InferenceClient task methods.
Route providers intentionally
Append :fastest, :cheapest, :preferred, or a named provider (e.g. :groq) to the model id. Default routing picks the fastest available provider.
Separate Hub seats from inference spend
Free, PRO, Team, and Enterprise cover Hub features and collaboration. Inference Providers and Endpoints bill separately as pay-as-you-go or dedicated hosting.
Load datasets efficiently
Use the datasets library with load_dataset. Enable streaming for large files. Preview rows in Data Studio before downloading.
Document models you publish
Upload a Model Card with task, training summary, evaluation, limitations, and license. Gated models require user acceptance before access.
Explore Spaces for quick demos
Gradio and Streamlit Spaces run in the browser. ZeroGPU adds on-demand NVIDIA GPUs for eligible demos.
Use structured chat messages
Set a system message for output format and constraints. Keep user messages focused on the task and source material.
Enable streaming for long replies
Set stream=True in chat completion calls to receive incremental tokens—useful for responsive UIs and early cancellation.
Check licenses before production
Model and dataset licenses vary (open, research-only, commercial restrictions). Confirm terms in the card before shipping.
Copy snippets from the widget
View Code Snippets on the model page generates Python or JavaScript that matches the widget endpoint—reduces integration errors.
Request provider support when needed
If a model has no widget, click Ask for provider support on the model page. Providers choose which models to host.
Know when to use Endpoints
Inference Endpoints provide dedicated, always-on deployment with autoscaling. Inference Providers are serverless and pay-per-call.
Pricing, features, limits and latest updates
Models, datasets, and inference for open machine learning.
Free / Paid · Free
Free
Free
PRO
$9/ Monthly
Team
$20/ Monthly
Enterprise
$50/ Monthly
Enterprise Plus
Custom
HuggingChat and the Inference Playground let you try models in a chat UI; API and widget availability depends on the model and provider.
Inference Providers and Endpoints are pay-as-you-go on top of Free, PRO, Team, and Enterprise.
Hosts community models, datasets, and Spaces; Hub seats are separate from GPU bills.
Share a prompt that worked for you. Username and email are shown with your submission. External links are not allowed.
Example prompt
Prompt
Open the model card for [model-id] on Hugging Face Hub. Read the Model Card limitations, then run one short prompt in the hosted widget. Compare the widget output to the card's intended use case.
Short explanation
Hugging Face workflows differ by surface: Hub widgets, Inference Providers API, and Spaces are separate paths.
Example prompt
Prompt
Using Inference Providers, send one chat completion to [provider/model] with: System: You are a concise assistant. User: Explain LoRA fine-tuning in three sentences for a backend engineer.
Short explanation
API usage is billed separately from Hub PRO seats; check Inference Providers credits and rate limits in official docs.
Required fields are marked. Variation, result, and explanation are optional.
AI coding agent for building software in your editor, terminal, and IDE.
CodingFree / Paid
Platform for building and running AI agents with visual Agent Builder, AutoPilot chat, scheduling, and 200+ integration blocks; self-host open source or use hosted Pro/Max plans.
CodingFree / Paid
Build full-stack web apps from prompts with AI-generated code, preview, and deploy.
CodingFree / Paid