Gradio 3.0 is Out!
Hugging Face
Hugging Face just shipped Gradio 3.0, a full rewrite of its ML demo library. New frontend, new Blocks tool for custom apps, plus a contest to build demos.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Gradio has quietly become the default way people show off machine learning models without making anyone touch a command line. Since 2019 the library has powered more than 600,000 demos, and today Hugging Face pushed out version 3.0, which is less an update and more a rebuild from the ground up.
The most visible change is the frontend. Gradio switched to Svelte, and the payoff shows up immediately: smaller file sizes, faster load times, and a cleaner look that fits better when demos get embedded inside blog posts or docs rather than living in their own isolated tab. The Dataframe component now accepts drag-and-drop CSV uploads, there's a new Gallery component for image-heavy outputs, and a TabbedInterface class lets developers bundle several related demos into one app instead of scattering them across separate links.
The bigger story, though, is Blocks. It's a new low-level API that gives developers actual control over layout and data flow, something the original Gradio Interface never allowed. Want outputs arranged in a custom grid instead of the default input-left, output-right split? Want the output of one model feeding directly into a second model, or a dropdown that changes its own options based on what a user typed earlier? Blocks handles all of that, and it's still just Python — no JavaScript required. Hugging Face's own example shows a two-tab app, one tab reversing text and the other flipping an image left to right, built in about twenty lines of code.
To push adoption, Hugging Face is running something called the Gradio Blocks Party through the end of May, essentially a demo-building contest with merch and prizes for the best submissions. It's a low-stakes way to get more people experimenting with the new API before it becomes the expected way to build anything beyond a basic single-input, single-output demo.
My take — AI-written commentary, not fact-checked reporting
This is the right move — demos are how normal people actually judge whether an AI system works, not benchmark tables nobody outside a lab reads. Betting on Svelte for speed and adding Blocks for real layout control shows Hugging Face understands that developer experience is the actual moat here, not model weights. The Blocks Party is a smart, cheap way to seed an ecosystem before people default to something clunkier.
Read more about this at: Hugging Face