Generates working user interfaces the way image models generate pictures

You write what the screen should show, in plain words, and it returns the screen as a short program: one line per part (a header, a bar chart, a table), wired together. A renderer turns that program into a live interface. It is built for apps on OpenUI, an open toolkit where an assistant answers with an interface instead of a paragraph. Give it the list of components in your own product and it builds screens from your parts, so the result already matches your design.

The difference is how it writes. A chat model produces code left to right, token by token. This one works like an image generator: it starts from noise on a 256-token canvas and locks in each piece as soon as it is sure, so a simple screen lands in about a second on one GPU and a dense dashboard in three to six.

Under the hood it is a finetune of Google's DiffusionGemma, 26B parameters with 4B active per step. On Thesys's own 46-brief screen test it scores 71.7 percent against 13 percent for the untouched base model. That is the maker's number, but the raw outputs are published so anyone can rescore them.

Open the demo Space, which runs on a free GPU. Pick the "product analytics dashboard" example, watch the code condense out of noise in the left pane and the finished dashboard appear in the preview on the right. Then edit the brief: add a date filter or a second table and check whether the rest of the layout stays intact. To run it yourself, one vLLM command serves it as an OpenAI-style chat endpoint, but the weights need roughly 26 GB of GPU memory even compressed, so this is a data-center card, not a laptop. The OpenUI renderers for React, Vue and Svelte turn the output into a page.

It only writes OpenUI's screen language, so you need the OpenUI renderer to see anything; it is not a general chat or coding model. There is no seed control, so the same brief can come back with a different layout each time. The screen test it wins is Thesys's own, and the model is ten days old, so expect odd component choices on briefs far from its training set.

• Turn a feature request into a clickable mockup before anyone opens a design tool.
• Let an internal assistant answer "how did sales do this week" with a chart and a table instead of text.
• Generate admin and settings screens from your existing component library so they match the product.
• Spin up a dozen dashboard layouts for a client in a minute and pick one to refine.

Try the demo

View model page

Read this issue on neuralexpedition.com →