Hugging Face UX review: what a synthetic audit found

Published September 1, 2026

Hugging FaceUX auditmachine learning platformshomepage UXsynthetic user testing

Hugging Face is a platform where developers publish and download machine learning models, datasets, and demo apps. It was in the news this week on reports that Nvidia is in talks to acquire it. We ran a synthetic UX audit on the homepage that same week, sending a persona suited to the site, a computer science student new to machine learning, through the page while it thought out loud.

What our persona hit on the Hugging Face homepage

• The hero splits attention. The left half carries the headline, subheadline, and two CTAs. The right half is a live catalog panel with Tasks, Parameters, Libraries, Apps, and Inference Providers filters, plus a scrolling list of model cards with timestamps and stats. The panel matches the subheadline's text size and weight and contains interactive controls, so it reads as a second primary element rather than as support for the first. Heuristic: NN/g heuristic 8, aesthetic and minimalist design.

• The two CTAs assume vocabulary the visitor may not have. They read "Explore AI Apps" and "Browse 2M+ models", separated by the word "or" and styled identically as outlined buttons. Nothing signals which is the gentler entry point. A student in a first machine learning course may not yet know that apps are user-facing tools built on models. Heuristic: NN/g heuristic 2, match between the system and the real world.

• Two visual languages with no bridge between them. Items in the Spaces column use bold gradient backgrounds and emoji. Items in the Models column are plain text links with metadata: update timestamps, download counts abbreviated as "159k", heart counts like "4.52k". Each column is internally consistent, but nothing on the page explains that Spaces are runnable apps while Models are downloadable weights. Heuristic: NN/g heuristic 2, match between the system and the real world.

• The three columns run together. Models, Spaces, and Datasets sit side by side with category headers and small icons, but no dividers or background tints separate them. The gradient cards in the Spaces column pull the eye horizontally across all three columns, working against the vertical grouping each category depends on. Heuristic: Gestalt, proximity and grouping.

One thing Hugging Face does well

Every trending item carries engagement metrics and a recency timestamp such as "Updated 5 days ago", which signals an active community at a glance. Each column then ends with a scale anchor: "Browse 2M+ models", "Browse 1M+ applications", "Browse 500k+ datasets". Together these frame the trending items as curated picks from a large, active collection rather than as the entire inventory.