Quality, measured

Accessibility

The library does not claim a conformance level: that is signed off by an external audit of a specific application, and none exists yet. What is public is what runs on every change:

In three engines

The behaviour tests (focus, overlays, keyboard, dragging, printing, motion) run in Chromium, and also in Firefox and WebKit: that is where two bugs that Chromium did not show came up (the focus Safari lost when closing a dropdown, and dialogs that vanished abruptly in Firefox and Safari). CI has a fast gate with both.

Size

Measured on October 19, 2026 with the benchmark in the library's repository: the same bundler (Vite, in production) and the same settings for all of them, with React left out of the count. It is the weight in the bundle, gzip-compressed, of what is requested and used:

And the same screen (a form with validation, an orders table with sorting and pagination, and a confirmation dialog) built with four libraries, rendered on the server and hydrated in Chromium with the CPU throttled four times (median of 15 loads):

• It is not a feature comparison. The @amset/ui Table sorts, filters, groups, virtualises and edits in cell; the MUI and Mantine ones do none of that on their own. What is measured is size.
• The CSS of HeroUI and shadcn is not included in the individual pieces: Tailwind generates it in each application. That is why those two, and react-aria-components, which ships no styles, have no equivalent screen. shadcn is not a library: what is measured is what its pieces pull in.
• MUI and antd put the CSS inside the JavaScript, and their HTML is heavier because they render it inside.
• The Mantine and antd screens change their theme: with the default one they fail AA contrast. It is corrected with their own settings.
• Hydration time depends on the machine: it is valid for comparing the libraries with each other in the same run, not as an absolute figure. The machine was not idle; that is why the median and the range are given. The sizes do not depend on that.
• Not measured: build time, memory, performance when typing or sorting thousands of rows, or the size with other bundlers.

Lightweight by default

• One file per component and sideEffects limited to .css files: the bundler takes the JavaScript and CSS of what you import, nothing more.
• Each component imports its own stylesheet: there is no global stylesheet loading the whole library on every page.
• Languages are loaded on demand: with seven configured, one is downloaded.
• Flags are split by region: Flag requests the drawings for the region of the flag it renders, not those of the whole world.
• Icons are imported one by one: there are 2,134 and only the ones you use add weight.
• Overlays that are not open do not mount their content: a closed date picker does not render its calendar.

Server, client and hydration

Components without interaction can be used from a server component; those with interaction carry 'use client' as far down as possible. No component reads "today" or "now" while rendering, and the spaces Intl writes are normalised: that way the server HTML and the browser HTML match even when they are rendered on different days, in different time zones or with different ICU versions. One test renders each component page on one day and hydrates it on another.

Security

What is up to you: check addresses that come from a user before using them as href, and sanitise the HTML from an editor on the server. A vulnerability is reported with a confidential issue in the amset/amset-ui project on gitlab.amset.tech, never in a public one. Only the latest minor version receives fixes.

Licence

MIT: it can be used in personal and commercial projects, modified and redistributed, keeping the copyright notice.