beer-sakthai/sak-family-agent/sakthai-hf-hub-models-widgets

Hugging Face Hub Inference Widgets

Overview

The model page inference widget allows users to test models directly on the Hub without writing any code. Widgets are powered by the Inference Providers network — a multi-provider inference serving system that replaced the old "Hosted Inference API."

Widget Configuration

Widgets are configured through YAML metadata in the model card:

widget:
  - text: "The two men running to become New York City's next mayor..."
    example_title: "Reading comprehension"
  - src: sample1.flac
    example_title: "Speech sample 1"

Example Input Types

• Text tasks: text property
• Audio/Image tasks: src property (URL or relative repo path)
• Multiple examples: Array under widget: key
• Examples repo-relative: Just use filename/path inside the repo
• Default inputs: Source at huggingface.js/packages/tasks/src/default-widget-inputs.ts

Example Outputs

For models without live provider support, example outputs can be specified:

widget:
  - text: "I liked this movie"
    output:
      - label: POSITIVE
        score: 0.8
      - label: NEGATIVE
        score: 0.2

Output types:

• Text output: output.text
• Label scores: output[].label + output[].score
• Media output: output.url (image/audio file in repo)

Widget Availability Conditions

1. Task Support: The model's pipeline tag must be served by at least one

Inference Provider

1. Provider Availability: At least one provider actively serves that model
2. Model Configuration: Proper metadata and config files must exist

If no provider supports a model, "Ask for provider support" button appears.

Inference Playground

The Inference Playground lets users test chat completion models interactively with custom prompts, compare models, and tweak parameters (temperature, max tokens). Uses the same Inference Providers infrastructure.

How the Widget Differs from InferenceClient

Key Takeaway

Widgets are the public face of Inference Providers on model pages. Configure them well = better user engagement. For zero-cost production inference, use InferenceClient with community-tier providers targeting the same model ID.